Datasets:
slug stringlengths 2 43 | name stringlengths 2 47 | type stringclasses 4
values | description stringlengths 118 599 | organization_slug stringlengths 3 39 | organization stringlengths 3 77 ⌀ | steward stringclasses 6
values | category_slug stringclasses 35
values | category stringclasses 35
values | layer stringclasses 3
values | group stringclasses 10
values | category_position int32 0 74 | openness_score int32 0 5 | openness_class stringclasses 11
values | openness_bucket stringclasses 3
values | openness_components stringlengths 33 1.18k | openness_confidence stringclasses 3
values | openness_note stringlengths 34 580 | openness_last_verified date32 | adoption_level int32 1 5 ⌀ | adoption_reach stringclasses 16
values | adoption_signal_type stringclasses 5
values | adoption_confidence stringclasses 3
values | adoption_note stringlengths 39 483 | adoption_last_verified date32 | capability_score int32 1 5 ⌀ | capability_basis stringclasses 5
values | capability_basis_detail stringlengths 4 428 ⌀ | capability_value stringlengths 40 813 ⌀ | capability_relative_to stringlengths 2 26 ⌀ | capability_relation stringclasses 4
values | capability_confidence stringclasses 3
values | capability_note stringlengths 31 595 | capability_last_verified date32 | overall_score float64 1 5 ⌀ | overall_basis stringclasses 2
values | tier stringclasses 2
values | mature bool 2
classes | languages listlengths 1 1.81k ⌀ | aliases listlengths 1 5 ⌀ | version_note stringlengths 30 942 ⌀ | end_of_life_date date32 | end_of_life_source stringclasses 3
values | lineage_derived_from listlengths 1 8 ⌀ | lineage_curated_with listlengths 1 4 ⌀ | lineage_trains listlengths 1 6 ⌀ | lineage_part_of listlengths | github listlengths 1 9 ⌀ | pypi listlengths 1 4 ⌀ | npm listlengths 1 3 ⌀ | crates listlengths 1 1 ⌀ | go listlengths | huggingface_model listlengths 1 54 ⌀ | huggingface_dataset listlengths 1 16 ⌀ | arxiv listlengths 1 3 ⌀ | citation_count int32 3 43 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
llama | Llama 4 Scout / Maverick | model | Llama 4 is Meta's natively multimodal mixture-of-experts family, released in April 2025. Scout has 17B active and 109B total parameters with a 10M-token context, and Maverick has 17B active and 400B total parameters with a 1M-token context. The larger Behemoth model has not been released. | meta | Meta | null | base_pretrained | Base / pretrained models | model_components | Models | 0 | 3 | open_weights | open-ish | weights:open(downloadable on HF);data:closed;code:partial(inference/utils via llama-models, no training code);license:Llama-4-Community(non-OSI - 700M MAU commercial cap, acceptable-use policy with an EU-domicile exclusion for the multimodal models, Llama naming requirement for derivatives) | high | Llama 4's weights download from Hugging Face once a user accepts Meta's Llama 4 Community License and submits a contact form. The license caps commercial use at 700 million monthly active users and requires derivatives to carry the Llama name, and its acceptable-use policy withholds the multimodal models from individua... | 2026-11-12 | 3 | 100K-1M | usage_volume | medium | Adoption is measured as Hugging Face downloads summed across the five declared artifacts spanning Llama 3.1 and Llama 4. The older Llama-3.1-8B checkpoint still accounts for most of that reach, even though the capability reading is taken on Llama 4 Maverick -- reach and capability sit on different releases of the same ... | 2026-11-08 | 3 | benchmark | Artificial Analysis Intelligence Index | Llama 4 Maverick - AA Intelligence Index 10 (estimated, v4.3.2), ranked #25/45, below average among open weight non-reasoning models of similar size (median 12); Llama 4 Scout 8 on the same index; Behemoth is unreleased. Index v5.0.1 scores no Llama model. Pretraining-suite readings: on the Apertus-70B card's six-task ... | null | null | medium | Llama 4 Maverick, the strongest released member of the family, scored 10 on version 4.3.2 of Artificial Analysis's Intelligence Index, below the median of 12 for open-weight non-reasoning models of similar size. Llama 4 Scout scored lower, and the larger Behemoth model was never released. Version 5 of the index, introd... | null | 3 | adoption_and_capability | null | false | null | [
"llama-3-1",
"llama-4-scout-maverick"
] | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/meta-llama/Llama-3.1-405B",
"https://huggingface.co/meta-llama/Llama-3.1-70B",
"https://huggingface.co/meta-llama/Llama-3.1-8B",
"https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E",
"https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E"
] | null | null | 18 |
deepseek | DeepSeek-V4-Pro | model | DeepSeek-V4-Pro is DeepSeek's 1.6T-parameter mixture-of-experts model, with 49B parameters active per token, a one-million-token context and a hybrid attention design that combines Compressed Sparse Attention with Heavily Compressed Attention. It was pretrained on more than 32T tokens and ships as base and instruct che... | deepseek | DeepSeek | null | base_pretrained | Base / pretrained models | model_components | Models | 1 | 3 | open_weights | open-ish | weights:open(MIT, on HF, both Pro and Flash);data:closed;code:partial(inference/serving; no training pipeline or data);license:MIT(permissive, no use restrictions) | high | DeepSeek releases the V4 line's weights under a plain MIT license, but publishes no training corpus and only inference and serving code rather than a full training pipeline. This makes DeepSeek an open-weights release rather than a fully open model. | 2026-11-12 | 4 | 1M-10M | usage_volume | high | Adoption is measured as Hugging Face downloads summed across the family's six shipped checkpoints, spanning the V3 line and the newer V4-Pro and V4-Flash releases. V4-Flash draws the largest share of that traffic, with the base checkpoints seeing comparatively little use. | 2026-11-08 | 5 | benchmark | Artificial Analysis Intelligence Index v5.0.1 | Artificial Analysis Intelligence Index v5.0.1, read 2026-11-12: DeepSeek V4.1 Flash 1017 (Max reasoning effort) 33, second of 27 large open-weight models (class median 23); DeepSeek V4 Pro 0813 (Max) 23, 16th of 27. Read on the same version and date: Kimi K3.1 Pro 46, GLM-5.3 (Max) 32, MiMo-V2.6-Pro 30, MiniMax-M3.1-Fl... | null | null | high | DeepSeek V4.1 Flash 1017, an October 2026 update to V4.1 Flash, is the family's strongest model on version 5.0.1 of Artificial Analysis's Intelligence Index. At maximum reasoning effort it places second among large open-weight models, behind Kimi K3.1 Pro and level with GLM-5.3 and MiMo-V2.6-Pro, the other models this ... | 2026-11-12 | 4.7 | adoption_and_capability | leading | false | null | [
"deepseek-v3-2",
"deepseek-v3-base",
"deepseek-v4-pro",
"deepseek-v4-pro-base"
] | Openness follows the V4-Pro line, while capability is read on the family's highest-scoring releases and adoption across the V3 and V4 checkpoints this entry tracks. | null | null | null | null | null | null | [
"https://github.com/deepseek-ai/DeepSeek-V3"
] | null | null | null | null | [
"https://huggingface.co/deepseek-ai/DeepSeek-V3.2",
"https://huggingface.co/deepseek-ai/DeepSeek-V3-Base",
"https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro",
"https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash",
"https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-Base",
"https://huggingface.co/deepse... | null | null | 19 |
qwen | Qwen Base | model | Alibaba's pretrained Qwen base models, published for fine-tuning and research rather than chat. The current base generation is Qwen3.5, released in February 2026 in five sizes from 0.8B to a 35B mixture of experts with 3B active parameters. Qwen has published no base checkpoints for Qwen 3.6 or 3.8, whose open-weight m... | alibaba-cloud | Alibaba Cloud | null | base_pretrained | Base / pretrained models | model_components | Models | 2 | 3 | open_weights | open-ish | weights:open(ungated on the Hub: the five Qwen3.5 Base checkpoints, 0.8B to 35B-A3B);data:closed(no corpus or data card published; the cards name no datasets and the series README describes the corpus only as trillions of multimodal tokens);code:partial(usage, deployment and fine-tuning pointers in QwenLM/Qwen3.8, whic... | high | Qwen3.5's base models download from Hugging Face under Apache-2.0 without a gate, in five sizes. The pretraining corpus is not released and only usage code is published, so this is open weights rather than open source. | 2026-11-12 | 4 | 1M-10M | usage_volume | high | Adoption is measured as Hugging Face downloads summed across the seven declared base checkpoints, the five Qwen3.5 Base sizes and the Qwen2.5 72B and 7B bases. The Qwen2.5-7B base still draws the most downloads of the seven. | 2026-11-08 | 4 | benchmark | Artificial Analysis Intelligence Index | Qwen3.5 35B A3B (Reasoning), the post-trained sibling of the largest Qwen3.5 Base checkpoint - AA Intelligence Index 19 (estimated, v4.3.2), well above the median of 8 for open-weight models of similar size. Qwen3.5 397B A17B scores 18. On the same index version Llama 4 Maverick (llama, 3) scores 10 and DeepSeek-V4.1-F... | null | null | medium | Artificial Analysis does not evaluate base checkpoints, so this record is read on its post-trained sibling, as llama is read on Llama 4 Maverick. Qwen3.5 35B A3B, the chat model built on the largest Qwen3.5 Base checkpoint, scored 19 on version 4.3.2 of Artificial Analysis's Intelligence Index, well above the median of... | null | 4 | adoption_and_capability | strong | false | null | [
"qwen-2-5",
"qwen3"
] | Covers Qwen's base checkpoints only. The post-trained models are in qwen-instruct, the same split as llama and llama-instruct; the aliases for the Qwen 3.6 to 3.8 releases point here from before that split. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/Qwen/Qwen2.5-72B",
"https://huggingface.co/Qwen/Qwen2.5-7B",
"https://huggingface.co/Qwen/Qwen3.5-35B-A3B-Base",
"https://huggingface.co/Qwen/Qwen3.5-9B-Base",
"https://huggingface.co/Qwen/Qwen3.5-4B-Base",
"https://huggingface.co/Qwen/Qwen3.5-2B-Base",
"https://huggingface.co/Qw... | null | null | 18 |
mistral-large | Mistral Large 3 | model | Mistral Large 3 is Mistral AI's first mixture-of-experts model since Mixtral, with 41B active and 675B total parameters, a 256K context window and native text-and-image input. Trained from scratch on 3,000 H200 GPUs and released in December 2025 in base and instruct form, it sits alongside the smaller dense Ministral 3... | mistral-ai | Mistral AI | null | base_pretrained | Base / pretrained models | model_components | Models | 3 | 3 | open_weights | open-ish | weights:open(Apache-2.0);data:closed;code:closed;license:Apache-2.0(OSI) | high | Mistral Large 3 ships Apache-2.0 weights, but the training data is closed and no training code is published. Mistral relicensed between releases: Large 2 shipped under the Mistral Research License, permitting research and non-commercial use only, and Large 3, on which this entry is scored, is Apache-2.0. Mistral Large ... | null | 1 | <10K | usage_volume | medium | Adoption is measured as Hugging Face downloads summed across the three declared artifacts, spanning the retired Mistral Large 2 and the Large 3 base and instruct checkpoints. The flagship is mostly consumed through Mistral's API rather than downloaded, so this understates real use, but only the artifacts a product decl... | null | 4 | benchmark | Artificial Analysis Intelligence Index v5.0.1 | Artificial Analysis Intelligence Index v5.0.1, read 2026-11-12: Mistral Large 4 (High reasoning effort) 26, 12th of 27 large open-weight models (class median 23). Mistral Large 3 has no v5.0.1 score. Read on the same version and date: MiMo-V2.6-Pro 30, MiniMax-M3.1-Flash 29, DeepSeek V4 Pro 0813 (Max) 23, Inkling (Xhig... | mimo-pro | one_below | high | Mistral Large 4, the 1-trillion-parameter model Mistral released in September 2026, is the family's strongest model on version 5.0.1 of Artificial Analysis's Intelligence Index. At high reasoning effort it scores above the median for large open-weight models and above DeepSeek V4 Pro, but a step below MiMo-V2.6-Pro and... | 2026-11-12 | 3.1 | adoption_and_capability | null | false | null | [
"mistral-large-2",
"mistral-large-3"
] | Capability is read on Mistral Large 4, published on Hugging Face in September 2026 under Mistral's Ecosystem Open-Weight License, as the family's strongest release. Openness and adoption are read on Large 3, the release this entry declares. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/mistralai/Mistral-Large-Instruct-2407",
"https://huggingface.co/mistralai/Mistral-Large-3-675B-Base-2512",
"https://huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512"
] | null | null | 11 |
gemma | Gemma 4 | model | Gemma 4 is Google DeepMind's open-weight model family, built from the same research as Gemini 3. It spans E2B and E4B edge models, an encoder-free 12B unified model, a 26B-A4B mixture-of-experts model and a 31B dense model, all taking text and image input, with audio on the three smallest. Context windows reach 256K to... | google | Google | null | base_pretrained | Base / pretrained models | model_components | Models | 4 | 3 | open_weights | open-ish | weights:open(Apache-2.0);data:closed;code:closed;license:Apache-2.0(OSI,but no open data/code) | medium | Gemma 4 ships under Apache-2.0 and is ungated on Hugging Face, unlike Gemma 2 and 3, which remain gated under the Gemma License and its Prohibited Use Policy. The model card describes the pretraining mixture without releasing it and links no training pipeline, so Gemma 4 is an open-weights release rather than a fully o... | 2026-11-12 | 5 | >10M | usage_volume | medium | Adoption is measured as Hugging Face downloads summed across six declared checkpoints spanning three generations. Almost all of that traffic now goes to the two Gemma 4 checkpoints, with the Gemma 2 and 3 repositories seeing comparatively little use. | 2026-11-08 | 4 | benchmark | Artificial Analysis Intelligence Index; model-card evaluation table | Artificial Analysis Intelligence Index v5.0.1, read 2026-11-12: Gemma 4 31B (Reasoning) 4, ninth of ten small open-weight models (class median 8). Read on the same version and date: Granite 4.2 8B 5, Inkling (Xhigh) 12, Mistral Large 4 (High) 26. gemma-4-31B-it model card - MMLU Pro 85.2%, GPQA Diamond 84.3%, AIME 2026... | null | null | medium | Gemma 4 31B, the family's strongest model, scores 4 on version 5.0.1 of Artificial Analysis's Intelligence Index, just below Granite 4.2 8B at 5 and well behind Inkling at 12 and Mistral Large 4 at 26. Gemma leads Granite on five of the index's nine evaluations and trails it only on AA-Omniscience, which penalizes wron... | null | 4.3 | adoption_and_capability | strong | false | null | [
"gemma-2",
"gemma-3",
"gemma-4"
] | Openness follows Gemma 4, while adoption is read across the Gemma 2, Gemma 3 and Gemma 4 checkpoints this entry tracks. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/google/gemma-2-27b",
"https://huggingface.co/google/gemma-2-9b",
"https://huggingface.co/google/gemma-3-27b-pt",
"https://huggingface.co/google/gemma-3-12b-pt",
"https://huggingface.co/google/gemma-4-31B-it",
"https://huggingface.co/google/gemma-4-26B-A4B-it"
] | null | null | 21 |
yi | Yi-1.5 | model | 01.AI's foundation model family at 6B, 9B and 34B with 32K-context variants, open sourced in May 2024. It continues pretraining from the original Yi release with a further 500 billion tokens, for 3.6 trillion in total. | 01-ai | 01.AI | null | base_pretrained | Base / pretrained models | model_components | Models | 5 | 3 | open_weights | open-ish | weights:open(Apache-2.0, on HF);data:closed(pretraining corpus not released);code:partial(inference/usage examples; no training pipeline);license:Apache-2.0(OSI, permissive, no use restrictions) | high | Permissive Apache-2.0 weights, but the training data and the full training code are not released, so this is open weights rather than open source. | 2026-11-12 | 1 | <10K | usage_volume | medium | Adoption is measured as Hugging Face downloads for the single declared checkpoint, 01-ai/Yi-1.5-34B. Yi is a 2024-era family now well behind the Qwen, Llama and DeepSeek generations, and its downloads have fallen off accordingly. | 2026-10-25 | 2 | benchmark | MMLU/BBH/C-Eval | Yi-34B base - MMLU 76.3, BBH 54.3, C-Eval 81.4. Competitive in the 2023-2024 mid-size class. | null | null | medium | Yi-34B, the original 2023 base model, scores 76.3 on MMLU, 54.3 on BBH and 81.4 on C-Eval in its tech report, competitive for its time but well below current DeepSeek and Qwen releases and comparable to Falcon 3. The Yi-1.5 model cards show base-model benchmarks only as charts, so the published figures are for Yi-34B r... | 2026-11-12 | 1.7 | adoption_and_capability | null | false | null | [
"yi-1-5"
] | Scored on the base models; the Chat releases are not separately scored. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/01-ai/Yi-1.5-34B"
] | null | null | 13 |
internlm | InternLM 2.5 | model | InternLM is Shanghai AI Lab's open-weights model family. The InternLM 2.5 generation, scored here on its 7B base model, handles a one-million-token context. Its successor InternLM3 is an 8B model trained on four trillion tokens, with a deep-thinking chain-of-thought mode beside normal responses. | shanghai-ai-laboratory | Shanghai AI Laboratory | null | base_pretrained | Base / pretrained models | model_components | Models | 6 | 3 | open_weights | open-ish | weights:open(custom 'Free Commercial License', academic free, commercial use free but requires an application form);data:closed(pretraining corpus not released);code:open(Apache-2.0 training/inference code on GitHub);license:Apache-2.0(code, OSI) + custom weights license(non-OSI, application step) | medium | The training and inference code is Apache-2.0, but the weights ship under a custom, non-OSI 'Free Commercial License': free for academic work and free for commercial use, but only once an application form has been approved. That application step on the weights, together with a pretraining corpus that is not released, i... | 2026-11-12 | 1 | <10K | usage_volume | medium | Adoption is measured as Hugging Face downloads for the single declared checkpoint, internlm2_5-7b. | 2026-11-08 | 2 | benchmark | MMLU | InternLM2.5-7B base: MMLU 71.6 (5-shot, OpenCompass); matches Yi-1.5-9B, beats Llama-3-8B (66.4). | null | null | medium | InternLM2.5-7B scores 71.6 on MMLU (5-shot, OpenCompass), matching Yi-1.5-9B and beating Llama-3-8B's 66.4, with a 1M-token context. That places it beside Falcon 3 and Yi-1.5 among small models from the same generation, well below the 2026 frontier. | 2026-11-12 | 1.7 | adoption_and_capability | null | false | null | [
"internlm-2-5"
] | Scored on the InternLM2.5-7B base checkpoint; InternLM3 has been released only as instruct models. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/internlm/internlm2_5-7b"
] | null | null | 9 |
gpt-4o | GPT-4o | model | GPT-4o is OpenAI's 2024 multimodal model, taking text and image input with a 128K context window and an October 2023 knowledge cutoff. It is still served through OpenAI's API, without released weights, but is deprecated there: its remaining snapshots are scheduled for removal on January 14, 2027, with GPT-6 Luna named ... | openai | OpenAI | null | base_pretrained | Base / pretrained models | model_components | Models | 7 | 1 | closed | closed | weights:closed;data:closed;code:closed;license:Proprietary(API-only) | high | GPT-4o is available only through the OpenAI API, with no downloadable weights. OpenAI publishes neither its training data nor its training code. | 2026-11-12 | null | null | unknown | low | GPT-4o is still served through the API until its announced removal, but OpenAI publishes no usage count for it, and ChatGPT's user figures no longer describe it: GPT-4o has left ChatGPT, which now runs on the GPT-6 family. No level is assigned until a count of GPT-4o's own use appears. | 2026-11-12 | 3 | benchmark | Same-harness benchmark comparison (OpenAI simple-evals, Phi-4 technical report) | Same-harness comparison in the Phi-4 technical report (OpenAI simple-evals, December 2024): GPT-4o MMLU 88.1, GPQA 50.6, MATH 74.6, HumanEval 90.6, MMLU-Pro 73.0, against Phi-4 (phi, 3) at 84.8, 56.1, 80.4, 82.6 and 70.4. Artificial Analysis Intelligence Index v5.0.1, read 2026-11-12: GPT-4o (Nov '24) has no score. | phi | at | medium | GPT-4o was OpenAI's multimodal flagship in 2024 and is still served through the API, but it sits far below the current frontier, and version 5 of Artificial Analysis's Intelligence Index does not score it. Microsoft's Phi-4 technical report ran both models through one evaluation harness: GPT-4o leads the 14-billion-par... | 2026-11-12 | null | null | null | false | null | null | null | 2027-01-14 | https://developers.openai.com/api/docs/deprecations | null | null | null | null | null | null | null | null | null | null | null | null | 9 |
reka-core | Reka Core | model | Reka Core is Reka AI's multimodal foundation model, announced in April 2024, natively processing text, images, video and audio with a 128K context window and delivered through an API or on-premises and on-device deployment. Reka has since published open weights for its smaller Flash and Edge models, but not for Core. | reka-ai | Reka AI | null | base_pretrained | Base / pretrained models | model_components | Models | 8 | 1 | closed | closed | weights:closed;data:closed(mix of public+licensed proprietary);code:closed;license:Proprietary(API/on-prem, no weights) | high | Reka launched Core as a proprietary multimodal model for API, on-premises and on-device deployment, with no downloadable weights. Its technical report describes training data drawn from public and proprietary or licensed datasets, and neither the corpus nor the training code is released. | 2026-11-12 | 2 | null | reported_traction | low | Reka Core is distributed through Reka's own API as well as Snowflake Cortex and Oracle Cloud Marketplace, but no usage has been disclosed through any of those channels, so the reading here is directional rather than measured. The Snowflake and Oracle listings are not established by the launch page cited here. | null | 2 | benchmark | MMMU | Comparable to GPT-4V on MMMU; outperforms Claude-3 Opus on multimodal human eval; surpasses Gemini Ultra on video (Apr 2024 claims) | null | null | low | Reka Core was frontier-class at its April 2024 launch, comparable to GPT-4V on MMMU and claiming wins over Claude 3 Opus on multimodal human evaluation and Gemini Ultra on video, though those comparisons are Reka's own launch claims rather than independent benchmarks. Two years on, that positioning is well behind the 2... | 2026-11-12 | 2 | adoption_and_capability | null | false | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | 6 |
inflection | Inflection 3.0 | model | Inflection AI's foundation model, powering the Pi assistant and the company's enterprise tier. Inflection 3.0, released October 2024, pairs a Pi conversational model with a Productivity model and runs on Intel Gaudi 3 hardware with on-premises or cloud deployment. Most of the original research team moved to Microsoft i... | inflection-ai | Inflection AI | null | base_pretrained | Base / pretrained models | model_components | Models | 9 | 1 | closed | closed | weights:closed;data:closed;code:closed;license:Proprietary(API/enterprise-licensed, no weights) | high | Inflection offers its models only through its API and enterprise licenses, with no downloadable weights. Models fine-tuned for a customer belong to that customer, but the base weights, training data and code stay with Inflection. | 2026-11-12 | 1 | null | reported_traction | low | After Microsoft's 2024 acqui-hire of most of the research team, Inflection pivoted to enterprise API access, but no usage of any kind -- users, customers or revenue -- has since been disclosed. With no published figure to read from, the reading here is a directional placement rather than one confirmed by a measured fig... | null | 2 | feature_matrix | null | null | null | null | low | Inflection 3.0 was built for emotional intelligence and productivity rather than leaderboard performance, and was never a frontier model even at release. Development halted after Microsoft took most of the research team, and no independent benchmark has appeared since, leaving it far behind the 2026 frontier models. | 2026-11-12 | 1.7 | adoption_and_capability | null | false | null | [
"inflection-3-0"
] | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | 8 |
glm | GLM-5.3 | model | GLM is Z.ai's open-weights model family, built for long-horizon agentic and coding tasks. GLM-5.3, the current release, takes a 1M-token context, and its predecessor GLM-5.2 is a mixture-of-experts model with 753B total and 40B active parameters. A smaller GLM-5.3-Flash ships beside it as a separately trained model. | zhipu-z-ai | Z.ai | null | base_pretrained | Base / pretrained models | model_components | Models | 10 | 3 | open_weights | open-ish | weights:open(GLM-5.3 License);data:closed;code:partial(inference and deployment code only);license:GLM-5.3-License(commercial use permitted, except that a model-as-a-service operator above USD 10 billion in annual revenue must first pass a Z.ai security review);sibling-sku:GLM-5.3-Flash(a smaller, separately trained mo... | high | GLM-5.3, the current release, ships its weights under the custom GLM-5.3 License. It reads like MIT with one bound: a company running a model-as-a-service business with more than USD 10 billion in annual revenue must pass a Z.ai security review before any commercial use. The training data is closed and only inference a... | 2026-11-12 | 4 | 1M-10M | usage_volume | high | Adoption is measured as Hugging Face downloads summed across four declared checkpoints spanning GLM-4 through GLM-5.2. GLM-5.2, the current release, accounts for nearly all of that total; the older THUDM/glm-4-9b repository now redirects to zai-org. | 2026-11-08 | 5 | benchmark | Artificial Analysis Intelligence Index v5.0.1 | Artificial Analysis Intelligence Index v5.0.1, read 2026-11-12: GLM-5.3 (Max reasoning effort) 32, third of 27 large open-weight models (class median 23). GLM-5.2 has no v5.0.1 score. Read on the same version and date: Kimi K3.1 Pro 46, DeepSeek V4.1 Flash 1017 (Max) 33, MiMo-V2.6-Pro 30. | null | null | high | GLM-5.3, the current release, places third among large open-weight models on version 5.0.1 of Artificial Analysis's Intelligence Index at maximum reasoning effort, just behind DeepSeek V4.1 Flash 1017 and ahead of MiMo-V2.6-Pro. Only Kimi K3.1 Pro scores clearly higher. Its predecessor GLM-5.2 has not been scored on th... | 2026-11-12 | 4.7 | adoption_and_capability | leading | false | null | [
"glm-4",
"glm-4-6",
"glm-4-9b",
"glm-5-1",
"glm-5-2"
] | Openness follows GLM-5.3, the release that currently governs. The GLM-5.3 checkpoints are not yet among the declared artifacts, so the adoption reading still sums GLM-4 through GLM-5.2. | null | null | null | null | null | null | [
"https://github.com/zai-org/GLM-5"
] | null | null | null | null | [
"https://huggingface.co/zai-org/GLM-4.6",
"https://huggingface.co/zai-org/glm-4-9b",
"https://huggingface.co/zai-org/GLM-5.1",
"https://huggingface.co/zai-org/GLM-5.2"
] | null | null | 13 |
kimi | Kimi K3 | model | Kimi K3 is Moonshot AI's flagship mixture-of-experts model, announced July 16, 2026, with 2.8 trillion total parameters activating sixteen of 896 experts, native vision and a one-million-token context. It is built on Kimi Delta Attention and attention residuals, and runs on kimi.com, Kimi Work, Kimi Code and the Kimi A... | moonshot-ai | Moonshot AI | null | base_pretrained | Base / pretrained models | model_components | Models | 11 | 3 | open_weights | open-ish | weights:open(shipped 2026-07-27, ungated, 96 safetensors shards);data:closed;code:partial(inference/deploy only; no training data or full pipeline);license:Kimi-K3-License(non-OSI; a model-as-a-service business with over USD 20M revenue in any 12 months needs a separate agreement before commercial use, and products abo... | high | Kimi K3's weights are ungated on Hugging Face under the Kimi K3 License, an MIT-style grant with two conditions. A company running a model-as-a-service business with more than $20 million in revenue over any 12 months must sign a separate agreement with Moonshot AI before commercial use, and a product with more than 10... | null | 4 | 1M-10M | usage_volume | high | Adoption is measured as Hugging Face downloads summed across the two declared artifacts, Kimi-K2.6 and Kimi-K3. K3's weights shipped on 2026-07-27, so the reading reflects downloads for the release that now governs the family rather than a longer production-traction record. | 2026-11-08 | 5 | benchmark | Artificial Analysis Intelligence Index v5.0.1 | Artificial Analysis Intelligence Index v5.0.1, read 2026-11-12: Kimi K3.1 Pro 46, first of 27 large open-weight models (class median 23), evaluated as a reasoning model with no effort variant listed. Kimi K3 has no v5.0.1 score. Read on the same version and date: DeepSeek V4.1 Flash 1017 (Max) 33, GLM-5.3 (Max) 32. | null | null | high | Kimi K3.1 Pro, post-trained from Kimi K3 and published in November 2026, is the highest-scoring open-weight model on version 5.0.1 of Artificial Analysis's Intelligence Index, well clear of DeepSeek V4.1 Flash 1017 and GLM-5.3. K3 itself has not been scored on this version. | 2026-11-12 | 4.7 | adoption_and_capability | leading | false | null | [
"kimi-k2-6",
"kimi-k3"
] | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/moonshotai/Kimi-K2.6",
"https://huggingface.co/moonshotai/Kimi-K3"
] | null | null | 9 |
inkling | Inkling | model | Inkling, released July 15, 2026, is Thinking Machines Lab's first open-weight foundation model, a natively multimodal mixture of experts with 975 billion total and 41 billion active parameters. Each token is routed to six of 256 experts plus two shared ones, with hybrid local and global attention, a one-million-token c... | thinking-machines-lab | Thinking Machines Lab | null | base_pretrained | Base / pretrained models | model_components | Models | 12 | 3 | open_weights | open-ish | weights:open(Apache-2.0 on HF, BF16 + NVFP4);data:closed(public+third-party+synthetic, not redistributed);code:partial(day-0 transformers/SGLang/vLLM/llama.cpp; no training pipeline);license:Apache-2.0(OSI; the weights)+Thinking-Machines-Model-AUP(conduct restrictions binding anyone who downloads the weights; no user, ... | high | Inkling's weights are released under Apache-2.0, but the training data and the full training pipeline are withheld, so it is an open-weights release rather than a fully open model. Downloading the weights also binds the user to Thinking Machines' Model Acceptable Use Policy, which restricts conduct such as weapons deve... | 2026-11-12 | 4 | 1M-10M | usage_volume | medium | Adoption is measured as Hugging Face downloads summed across the shipped checkpoints, the BF16 release and its NVFP4 quantization, so the family total is used rather than the headline repository alone. | null | 4 | benchmark | SWE-bench/GPQA/AIME | SWE-bench Verified 77.6%; SWE-bench Pro 54.3%; GPQA Diamond 87.2%; AIME 2026 97.1% (effort=0.99). Artificial Analysis Intelligence Index v5.0.1, read 2026-11-12: Inkling (Xhigh reasoning effort) 12, 24th of 27 large open-weight models (class median 23); Inkling Small 13, 22nd of 27. Read on the same version and date: M... | kimi | one_below | high | On its own evaluation table Inkling beats Nemotron 3 Ultra on SWE-bench Verified (77.6 vs 70.7) but trails Kimi K2.6 (80.2) and DeepSeek-V4-Pro (80.6), and Thinking Machines' launch post says Inkling is not the strongest overall model available, open or closed. Artificial Analysis's independent index places it lower: o... | null | 4 | adoption_and_capability | strong | false | null | null | The Inkling-Small sibling is folded into this record. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/thinkingmachines/Inkling",
"https://huggingface.co/thinkingmachines/Inkling-NVFP4"
] | null | null | 16 |
lucie-7b | Lucie 7B | model | Linagora and OpenLLM-France's fully open multilingual 7B foundation model on the Llama 3.1 architecture, trained on roughly three trillion tokens with about equal French and English shares plus German, Spanish, Italian and code. The weights ship with the Lucie Training Dataset, the Megatron-DeepSpeed training code and ... | linagora | Linagora | null | base_pretrained | Base / pretrained models | model_components | Models | 13 | 5 | open_source | open | weights:open(Apache-2.0);data:open(Lucie-Training-Dataset on HF, redistributed in training form);code:open(Lucie-Training Megatron-DeepSpeed fork);checkpoints:open(intermediate HF revisions);license:Apache-2.0(OSI) | high | Lucie-7B publishes Apache-2.0 weights with its full training stack: the Lucie-Training-Dataset corpus, the Megatron-DeepSpeed training code, and intermediate checkpoints as repository branches, matching OLMo and Apertus's openness tier. The corpus itself is CC-BY-NC-SA-4.0, published and ungated but not commercially re... | 2026-11-12 | 1 | <10K | usage_volume | medium | Adoption is measured as Hugging Face downloads summed across the base and instruct checkpoints. | 2026-11-08 | 2 | benchmark | multilingual-pretrain-suite | null | null | null | medium | Lucie-7B is a 7B multilingual base prioritizing French cultural representation and data-rights constraints, competitive for its size and language mix on French/English evaluations in the paper. It remains well below the 2026 open-weight frontier of trillion-parameter mixture-of-experts models. | 2026-11-12 | 1.7 | adoption_and_capability | null | false | null | null | Instruct variants are folded into this record; the successor Luciole family is a separate entry. | null | null | null | null | null | null | [
"https://github.com/OpenLLM-France/Lucie-Training"
] | null | null | null | null | [
"https://huggingface.co/OpenLLM-France/Lucie-7B",
"https://huggingface.co/OpenLLM-France/Lucie-7B-Instruct-v1.1"
] | null | null | 14 |
luciole | Luciole | model | Luciole is Linagora and OpenLLM-France's fully open French-English model family and the successor to Lucie, released at 1B, 8B and 23B in base and instruct forms. The weights ship with a 4.65-trillion-token training dataset drawn from openly licensed corpora, the training pipeline and training checkpoints. Base models ... | linagora | Linagora | null | base_pretrained | Base / pretrained models | model_components | Models | 14 | 5 | open_source | open | weights:open(Apache-2.0, safetensors on HF for 1B/8B/23B base + Instruct-1.1);data:open(Luciole-Training-Dataset on HF, ~4.65T tokens, "only corpora under open licenses");code:open(Luciole-Training pretraining + data-processing pipeline on GitHub, GPL-3.0);checkpoints:open(intermediate + final pretraining checkpoints r... | high | Luciole ships Apache-2.0 weights with its full training stack, including the CC-BY-SA-4.0 Luciole-Training-Dataset of openly licensed corpora, a GPL-3.0 training pipeline, and checkpoints from every training phase that the model card links on Linagora's download server. Unlike Lucie, whose whole corpus was non-commerci... | 2026-11-12 | 1 | <10K | usage_volume | medium | Adoption is measured as Hugging Face downloads summed across six checkpoints spanning the 1B, 8B and 23B sizes in base and instruct form. That total reflects a niche audience for a sovereign French-language model rather than broad usage. | 2026-11-08 | 2 | benchmark | multilingual-pretrain-suite | Luciole-23B-Base card charts (base models, one harness): English averaged normalized score about 49 against EuroLLM-22B-2512 about 48, Mistral-Small-24B-Base-2501 about 60 and Gaperon-1125-24B about 42; French about 42 against EuroLLM-22B about 34, Mistral-Small about 42 and Gaperon about 32. Values read from the chart... | null | null | low | Luciole is a French-English family up to 23B parameters built around French cultural representation and open data rights. Its 23B base model's card charts it against EuroLLM-22B and Mistral Small 24B on the same evaluations: about even with EuroLLM-22B in English and roughly 8 points ahead of it in French, where it als... | null | 1.7 | adoption_and_capability | null | false | null | null | null | null | null | null | null | null | null | [
"https://github.com/OpenLLM-France/Luciole-Training"
] | null | null | null | null | [
"https://huggingface.co/OpenLLM-France/Luciole-23B-Base",
"https://huggingface.co/OpenLLM-France/Luciole-23B-Instruct-1.1",
"https://huggingface.co/OpenLLM-France/Luciole-8B-Base",
"https://huggingface.co/OpenLLM-France/Luciole-8B-Instruct-1.1",
"https://huggingface.co/OpenLLM-France/Luciole-1B-Base",
"ht... | [
"https://huggingface.co/datasets/OpenLLM-France/Luciole-Training-Dataset"
] | null | 22 |
olmo | OLMo 3 | model | Ai2's OLMo 3 family at 7B and 32B in Base, Think, Instruct and RL-Zero variants, released November 2025 with a December update. The weights ship with the 9.3-trillion-token Dolma 3 corpus, the training code, the data recipes and intermediate checkpoints, so the whole run can be inspected and modified. It supersedes OLM... | allen-institute-for-ai | Allen Institute for AI | null | base_pretrained | Base / pretrained models | model_components | Models | 15 | 5 | open_source | open | weights:open(Apache-2.0);data:open(Dolma 3, 9.3T tokens, all stages);code:open(training code+recipes+logs);checkpoints:open(intermediate);license:Apache-2.0(OSI) | high | OLMo publishes Apache-2.0 weights, the full Dolma 3 training corpus, training code and intermediate checkpoints, all ungated and downloadable, so the whole training run is open to inspection. | 2026-11-12 | 3 | 100K-1M | usage_volume | medium | Adoption is measured as Hugging Face downloads summed across the two declared checkpoints, the 32B and 7B OLMo 3 releases. | 2026-11-08 | 4 | benchmark | MMLU/GPQA/reasoning | OLMo 3.1 Think 32B tops Qwen3-32B on AIME 2025 (78.1 vs 70.9) but trails it on GPQA (57.5 vs 67.3) and MMLU (86.4 vs 88.8) | null | null | high | OLMo 3.1 Think 32B beats Qwen3-32B on AIME 2025 but trails it on GPQA and MMLU, and the line as a whole is competitive with Qwen2.5 and Gemma 3. It remains a step below the trillion-parameter open-weight models at the frontier. | 2026-11-12 | 3.7 | adoption_and_capability | null | false | null | [
"olmo-3"
] | null | null | null | null | null | null | null | [
"https://github.com/allenai/OLMo"
] | null | null | null | null | [
"https://huggingface.co/allenai/Olmo-3-1125-32B",
"https://huggingface.co/allenai/Olmo-3-1025-7B"
] | null | null | 13 |
pythia | Pythia | model | EleutherAI's suite of sixteen models from 70M to 12B parameters, eight sizes trained twice on The Pile and its deduplicated variant, in the same data order, with 154 intermediate checkpoints released for each. It is a research tool for studying how language models learn rather than a production family, and the dataload... | eleutherai | EleutherAI | null | base_pretrained | Base / pretrained models | model_components | Models | 16 | 5 | open_source | open | weights:open(Apache-2.0);data:open(the Pile, with dataloader reconstruction tools);code:open(training+analysis code, GitHub);checkpoints:open(154 per model);license:Apache-2.0(OSI) | high | Pythia releases Apache-2.0 weights, the Pile as its training corpus together with dataloader tools that reconstruct the exact token order, training and analysis code, and 154 intermediate checkpoints per model size. Every part of the release is downloadable and the training run itself is reproducible, not just the mode... | 2026-11-12 | 1 | <10K | usage_volume | high | Adoption is measured as Hugging Face downloads for the single declared artifact, EleutherAI/pythia-12b. | null | 1 | benchmark | null | null | null | null | high | Pythia was never built to compete on capability. Its sixteen models top out at 12B, were trained on the 2020 Pile corpus, and are research tools for studying how language models learn rather than production systems. No benchmark measures it against the 2026 frontier, and it trails every model built for competitive perf... | 2026-11-12 | 1 | adoption_and_capability | null | false | null | null | null | null | null | null | null | null | null | [
"https://github.com/EleutherAI/pythia"
] | null | null | null | null | [
"https://huggingface.co/EleutherAI/pythia-12b"
] | null | null | 11 |
phi | Phi-4 | model | Phi-4 is Microsoft's 14B dense small language model, announced in December 2024 and fully released the following month. It was trained on 9.8 trillion tokens of heavily curated and synthetic data, with a recipe that centers on data quality and uses synthetic data throughout training. The line has since been extended wi... | microsoft | Microsoft | null | base_pretrained | Base / pretrained models | model_components | Models | 17 | 3 | open_weights | open-ish | weights:open(MIT);data:closed;code:closed;license:MIT(OSI) | high | Phi-4 ships MIT weights, genuinely permissive with no use restrictions. Neither the training corpus nor the training pipeline is published, however: Microsoft describes the synthetic data mix without releasing it, so the model is free to use but the recipe is not reproducible. | 2026-11-12 | 3 | 100K-1M | usage_volume | high | Adoption is measured as Hugging Face downloads for the single declared artifact, microsoft/phi-4. | 2026-11-08 | 3 | benchmark | GPQA | Phi-4 (14B) model card: GPQA 56.1 (graduate-level science) and MATH 80.4 (competition math); the technical report says it substantially surpasses its teacher, GPT-4, on STEM-focused question answering. Artificial Analysis Intelligence Index v5.0.1 does not score Phi-4. | null | null | high | Phi-4 is a 14B model with outsized math and reasoning results, including 56.1% on graduate-level science questions and 80.4% on competition math, and its technical report says it surpasses GPT-4, its teacher, on STEM question answering. It remains a small, older model against the 2026 frontier, and Artificial Analysis'... | 2026-11-12 | 3 | adoption_and_capability | null | false | null | [
"phi-4"
] | Scores base Phi-4 14B; the reasoning variants are separate SKUs. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/microsoft/phi-4"
] | null | null | 12 |
falcon | Falcon 3 | model | Falcon 3 is TII's open-weights model family from Abu Dhabi, spanning 1B to 10B parameters. It comes from a single large pretraining run on the 7B model over 14 trillion tokens of web, code, STEM, curated and multilingual data, extended by depth up-scaling for reasoning. TII has since released the Falcon-H1 family on th... | technology-innovation-institute | Technology Innovation Institute | null | base_pretrained | Base / pretrained models | model_components | Models | 18 | 3 | open_weights | open-ish | weights:open;data:closed;code:closed;license:TII-Falcon-License-2.0(Apache-based+AUP,non-OSI) | high | Falcon's weights are downloadable and free to use commercially at any scale under the TII Falcon License 2.0, an Apache-based license with an acceptable-use policy rather than a usage cap. TII does not release the training corpus or training code, and the acceptable-use clause keeps the license outside the OSI definiti... | 2026-11-12 | 1 | <10K | usage_volume | high | The single declared artifact, Falcon3-10B-Base, is scored on measured Hugging Face downloads rather than a reported-traction figure, since a direct count is the stronger signal when one exists. That count confirms what the Falcon 3 family's own description already says: relevance has faded against newer small language ... | 2026-11-08 | 2 | benchmark | MMLU-Pro | null | null | null | high | Falcon3-10B-Base scores 73.1 on MMLU and 42.5 on MMLU-Pro, and the 7B base 67.4 and 39.2. Those results were competitive among models under 13B when Falcon 3 was released, but sit well below the 2026 frontier. | 2026-11-12 | 1.7 | adoption_and_capability | null | false | null | [
"falcon-3"
] | The entry is scored on the Falcon 3 generation, now dated relative to the newer Falcon-H1 family. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/tiiuae/Falcon3-10B-Base"
] | null | null | 12 |
minimax | MiniMax M3 | model | MiniMax-M3 is MiniMax's flagship model, launched on June 1, 2026, with MiniMax Sparse Attention, a one-million-token context and native multimodality. It was API-first at launch, with weights subsequently published on Hugging Face alongside vLLM and SGLang deployment guides. | minimax | MiniMax | null | base_pretrained | Base / pretrained models | model_components | Models | 19 | 3 | open_weights | open-ish | weights:open(on HF; ungated; safetensors/BF16; MoE);data:closed;code:partial(inference only; no training pipeline/data);license:minimax-community(custom, see HF LICENSE) | high | MiniMax distributes MiniMax-M3's weights under a custom minimax-community license over a closed training corpus, with inference-only code and no training pipeline released. That combination of open weights with closed data and code makes MiniMax an open-weights release rather than a fully open model. | null | 3 | 100K-1M | usage_volume | high | Adoption is measured as trailing-30-day Hugging Face downloads summed across the two declared checkpoints, MiniMax-M3 and its successor MiniMax-M3.1-Flash. MiniMax is also served through OpenRouter and offered free inside coding tools such as Kilo Code, channels a download count does not capture. | null | 5 | benchmark | Artificial Analysis Intelligence Index v5.0.1 | Artificial Analysis Intelligence Index v5.0.1, read 2026-11-12: MiniMax-M3.1-Flash 29, sixth of 27 large open-weight models (class median 23), evaluated as a reasoning model with no effort variant listed; MiniMax-M3 19, 17th of 27. Read on the same version and date: GLM-5.3 (Max) 32, MiMo-V2.6-Pro 30, Mistral Large 4 (... | null | null | high | MiniMax-M3.1-Flash, the October 2026 successor built on MiniMax-M3 at the same 428-billion-parameter scale, scores 29 on version 5.0.1 of Artificial Analysis's Intelligence Index, sixth among large open-weight models and level with MiMo-V2.6-Pro. M3 itself scores 19, below the median for its class, so it is the newer r... | 2026-11-12 | 4.4 | adoption_and_capability | strong | false | null | [
"minimax-m2-5",
"minimax-m3"
] | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/MiniMaxAI/MiniMax-M3",
"https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash"
] | null | null | 8 |
apertus | Apertus | model | Switzerland's national open LLM family from the Swiss AI Initiative at EPFL, ETH Zurich and CSCS. The current 1.5 models at 8B and 70B take image and audio input alongside text, run a deliberation mode and handle a 262,144-token context. The line covers 1,811 languages, including Swiss German and Romansh, and about 40 ... | swiss-ai-initiative | Swiss AI Initiative | null | base_pretrained | Base / pretrained models | model_components | Models | 20 | 5 | open_source | open | weights:open(Apache-2.0, BF16 safetensors);data:open(full pretraining-data reconstruction scripts, github.com/swiss-ai/pretrain-data);code:open(Megatron-LM training recipe + reconstruction scripts);checkpoints:open(intermediate checkpoints on HF branches);license:Apache-2.0(OSI) | high | Apertus 1.0 publishes its pretraining-data reconstruction scripts, Megatron-LM training recipe, intermediate checkpoints and technical report under Apache-2.0, as OLMo and Pythia do. Its acceptable-use policy restricts conduct, not commercial use. Apertus 1.5, released 2026-07-24, continues training the 1.0 models unde... | 2026-11-12 | 2 | 10K-100K | usage_volume | medium | Adoption is measured as Hugging Face downloads summed across the family's three shipped checkpoints, most of them now for the multimodal v1.5 release that shipped 2026-07-24. The original 2509 flagship draws little of that traffic today, a shift from when it was the only checkpoint available. | null | 3 | benchmark | MMLU / pretraining suite | Apertus-70B pretraining-eval avg 67.5% (ARC/HellaSwag/WinoGrande/XNLI/XCOPA/PIQA), ~on par with Llama 3.1-70B (67.3%); Apertus-70B-Instruct MMLU 69.6%, trailing Llama 3.3-70B (87.5%) and Qwen2.5-72B (86.6%) by ~17-20 pts. | llama | at | medium | Apertus-70B averages 67.5% on pretraining benchmarks, matching Llama 3.1-70B's 67.3%, and is unusually strong multilingually. Its instruct model trails the newer frontier on reasoning, scoring 69.6% on MMLU against 87.5% for Llama 3.3-70B and 86.6% for Qwen2.5-72B. | null | 2.7 | adoption_and_capability | null | false | null | null | null | null | null | null | null | null | null | [
"https://github.com/swiss-ai/pretrain-code"
] | null | null | null | null | [
"https://huggingface.co/swiss-ai/Apertus-v1.5-70B",
"https://huggingface.co/swiss-ai/Apertus-v1.5-8B",
"https://huggingface.co/swiss-ai/Apertus-70B-2509"
] | null | null | 17 |
smollm | SmolLM3 | model | SmolLM3 is Hugging Face's fully open 3B small language model, released in base and instruct form in July 2025. Trained on about 11 trillion tokens, it supports a 128K context window, six languages and dual-mode reasoning. The full engineering blueprint, covering architecture, data mixtures and the training and post-tra... | hugging-face | Hugging Face | null | base_pretrained | Base / pretrained models | model_components | Models | 21 | 5 | open_source | open | weights:open(Apache-2.0);data:open(public data mixtures, ~11T tokens);code:open(training+post-training recipe published);checkpoints:open(SmolLM3-3B-checkpoints, every 40k steps);license:Apache-2.0(OSI) | high | SmolLM3 publishes Apache-2.0 weights alongside a complete engineering blueprint: architecture details, exact data mixtures and training methodology for its 11-trillion-token run, built entirely from public datasets and training frameworks. The Hugging Face card is Apache-2.0 licensed and ungated. | 2026-11-12 | 3 | 100K-1M | usage_volume | medium | Adoption is measured as Hugging Face downloads for the single declared checkpoint, SmolLM3-3B. It sees real on-device and edge use, but stays a niche choice next to the large open-weight families. | 2026-11-08 | 3 | benchmark | reasoning/multilingual | Outperforms Llama-3.2-3B and Qwen2.5-3B; competitive with 4B Qwen3/Gemma3 | null | null | high | SmolLM3-3B outperforms Llama-3.2-3B and Qwen2.5-3B and stays competitive with the larger 4B Qwen3 and Gemma 3. That comparison is against other small models rather than the broader frontier. | 2026-11-12 | 3 | adoption_and_capability | null | false | null | [
"smollm3"
] | null | null | null | null | null | null | null | [
"https://github.com/huggingface/smollm"
] | null | null | null | null | [
"https://huggingface.co/HuggingFaceTB/SmolLM3-3B"
] | null | null | 13 |
rwkv | RWKV | model | RWKV is an open-weights model family whose architecture runs as a recurrent network at inference yet trains in parallel like a transformer, giving linear-time inference, constant memory with no KV cache and effectively unbounded context. The RWKV-7 Goose generation, which includes the G1 models, adds dynamic state evol... | rwkv | RWKV Project | null | base_pretrained | Base / pretrained models | model_components | Models | 22 | 4 | open_weights | open-ish | weights:open(Apache-2.0);data:components-listed(itemized machine-readable component index with resolvable links to public datasets, but no released corpus and no reconstruction pipeline);code:open(github.com/BlinkDL/RWKV-LM);license:Apache-2.0(OSI) | high | RWKV publishes Apache-2.0 weights and open training code, plus a machine-readable index naming every training-data component with a resolvable link, category and citation. But the corpus itself is not released, mixing proportions are withheld, one component is gated and another's version unresolved, so the data stays i... | 2026-11-12 | 1 | <10K | usage_volume | medium | Adoption is measured as Hugging Face downloads summed across the two repositories RWKV declares, rather than GitHub stars, which had overstated the family's reach. Counting every repository under the BlinkDL and RWKV organizations would read higher, but only the artifacts a product declares count toward it. | 2026-11-08 | 2 | benchmark | MMLU on the RWKV-7 World model card, with the RWKV-7 Goose paper (arXiv:2503.14456) | RWKV-7 World 2.9B: MMLU 54.56% (1.5B: 44.84%) on the model card; the paper claims a new 3B-scale SoTA on multilingual tasks and parity with the 3B SoTA in English. The larger G1 models (up to 13.3B) publish no benchmark table. | internlm | at | medium | RWKV's case is efficiency rather than absolute capability: a linear-time recurrent architecture that its paper shows competitive at 3B scale on far fewer training tokens. Its best published MMLU sits well below InternLM2.5-7B and Yi-34B, small and mid-size models from the 2023-2024 generation that this map places at th... | 2026-11-12 | 1.7 | adoption_and_capability | null | false | null | null | null | null | null | null | null | null | null | [
"https://github.com/BlinkDL/RWKV-LM"
] | null | null | null | null | [
"https://huggingface.co/BlinkDL/rwkv-7-world",
"https://huggingface.co/RWKV/RWKV7-Goose-World3-1.5B-HF"
] | [
"https://huggingface.co/datasets/RWKV/RWKV-World-Listing",
"https://huggingface.co/datasets/Goose-World/RWKV-World-v3"
] | [
"https://arxiv.org/abs/2503.14456"
] | 12 |
ernie | ERNIE 5.1 | model | ERNIE 5.1 is Baidu's fifth-generation flagship language model, launched in May 2026 with stronger text capabilities than its predecessor. It is offered only as a hosted model, through Baidu's ERNIE site and the Baidu Qianfan platform. | baidu | Baidu | null | base_pretrained | Base / pretrained models | model_components | Models | 23 | 1 | closed | closed | weights:closed(no ERNIE 5.1 weights are distributed; the flagship is served only as a hosted model);data:closed;code:closed;license:proprietary(Baidu Qianfan API);API-only(ERNIE 4.5 was released openly, the 5.1 flagship is closed);separate-line:ERNIE 5.5 Flash Base, ERNIE 5.5 Flash Base Model License v1.0(a smaller 280... | high | Baidu distributes no weights for its flagship ERNIE 5.1, which is available only through its own services and API, and publishes neither its training data nor its code. ERNIE 5.5 Flash Base, which Baidu released separately on Hugging Face in October 2026, is a smaller open-weight base model under its own license; it is... | 2026-11-12 | 4 | null | reported_traction | low | The freshest available signal is Baidu's own app-level MAU for the ERNIE/Wenxiaoyan assistant, not a figure specific to ERNIE 5.1, and the models are not on Hugging Face, so there is no download count to read instead. Baidu's most recent quarterly release gives no ERNIE MAU at all, only a Baidu App total. | 2026-11-12 | 5 | benchmark | LMArena text and search leaderboards; Baidu-reported AIME 2026 | LMArena text leaderboard, snapshot of 2026-11-09: ERNIE 5.1 1467.7, 35th of 421, against MiMo-V2.6-Pro 1491.7, GLM-5.3 (Max) 1475.9, Kimi K3 (Max) 1474.3, DeepSeek-V4.1-Flash-20260910 (Max) 1463.4, Kimi K3.1 Pro (Max) 1457.3, DeepSeek-V4-Pro 1451.4 and MiniMax-M3 1428.2. LMArena search leaderboard: 1223.4 and 4th on 20... | null | null | medium | On LMArena's independent text leaderboard ERNIE 5.1 places just behind GLM-5.3 and Kimi K3 and level with DeepSeek's V4.1 Flash, and on the search leaderboard it posted 1,223 at its May 2026 launch, fourth globally, before newer models pushed it to eleventh. Baidu adds a 99.6 on AIME 2026 with tool use, but Artificial ... | null | 4.7 | adoption_and_capability | leading | false | null | [
"ernie-5-1"
] | Scored on Baidu's flagship line, currently the API-only ERNIE 5.1. ERNIE 5.5 Flash Base, an open-weight base model Baidu published separately on Hugging Face in October 2026, is a different line and is not scored here. ERNIE 4.5 was released openly. | null | null | null | null | null | null | null | null | null | null | null | null | null | null | 11 |
mimo-pro | MiMo-V2.6-Pro | model | MiMo-V2.6-Pro is Xiaomi's flagship mixture-of-experts model, with about 1.02 trillion total and 42 billion active parameters. Xiaomi publishes the weights under MIT, along with a fork of verl that reproduces its agentic RL recipe on a distilled 9B model, and a dataset of RL environments for code, cybersecurity, general... | xiaomi | Xiaomi | null | base_pretrained | Base / pretrained models | model_components | Models | 24 | 3 | open_weights | open-ish | weights:open(MiMo-V2.6-Pro-RL and MiMo-V2.6-Pro-MOPD on Hugging Face, ungated safetensors);data:closed(no pretraining corpus released or named);code:partial(inference support and RL post-training code (a verl fork with five environment recipes); no pretraining pipeline);license:MIT(HF card on both V2.6 Pro checkpoints)... | high | Xiaomi publishes the MiMo-V2.6-Pro weights under MIT, along with reinforcement-learning code and environments that reproduce its training recipe on a distilled 9B model. The pretraining corpus is not released and no pretraining code ships, so the model is open weights rather than fully open. | 2026-11-12 | 2 | 10K-100K | usage_volume | medium | Adoption is read from Hugging Face downloads of MiMo-V2.5-Pro, the only checkpoint declared when it was measured; the V2.6 checkpoints are newer than that reading. | 2026-11-08 | 5 | benchmark | Artificial Analysis Intelligence Index v5.0.1 | Artificial Analysis Intelligence Index v5.0.1, read 2026-11-12: MiMo-V2.6-Pro 30, fourth of 27 large open-weight models (class median 23), evaluated as a reasoning model with no effort variant listed. Read on the same version and date: Kimi K3.1 Pro 46, DeepSeek V4.1 Flash 1017 (Max) 33, GLM-5.3 (Max) 32. Kimi K3, the ... | null | null | high | MiMo-V2.6-Pro places fourth among large open-weight models on version 5.0.1 of Artificial Analysis's Intelligence Index, close behind DeepSeek V4.1 Flash 1017 and GLM-5.3 and well above the median for its class. That keeps Xiaomi's flagship among the strongest open-weight models available. | 2026-11-12 | 4.1 | adoption_and_capability | strong | false | null | [
"mimo-v2-5-pro",
"mimo-v2-6-pro"
] | Read on the Pro checkpoints. The Flash sibling and the distilled Qwen3.5-9B model are separate models and are not declared here. The older MiMo repository carries different terms from the model card, and the card is authoritative. | null | null | null | null | null | null | [
"https://github.com/XiaomiMiMo/MiMo",
"https://github.com/XiaomiMiMo/verl"
] | null | null | null | null | [
"https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro",
"https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL",
"https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-MOPD"
] | null | null | 11 |
alia-40b | ALIA-40b | model | ALIA-40b is a 40.4B-parameter decoder-only transformer pretrained from scratch by the Barcelona Supercomputing Center on the MareNostrum 5 supercomputer, funded by Spain's Ministry for Digital Transformation and the Civil Service. It was trained on 9.37 trillion tokens across 35 European languages and 92 programming la... | barcelona-supercomputing-center | Barcelona Supercomputing Center | null | base_pretrained | Base / pretrained models | model_components | Models | 25 | 4 | open_weights | open-ish | weights:open(Apache-2.0, safetensors on HF BSC-LT/ALIA-40b);data:components-listed(60+ named sources described in the model card, but the pretraining corpus "will not be released or distributed to third parties");code:open(all training scripts + configuration files at github.com/langtech-bsc/alia);checkpoints:open(part... | high | ALIA-40B's Apache-2.0 weights ship with the complete training code and recipe on GitHub, but the pretraining corpus is documented rather than released or reconstructable. That gap keeps it at open weights rather than the open-source tier that OLMo, Pythia and Apertus reach, the last by publishing data-reconstruction sc... | 2026-11-12 | 1 | <10K | usage_volume | medium | BSC-LT/ALIA-40b's only usage signal is Hugging Face downloads for the single declared checkpoint, and the intended audience is narrow: Spanish and Spain's co-official languages, plus the wider European-language set, rather than a global one. Downloads across the wider ALIA Kit would read higher, but only the base model... | 2026-11-08 | 2 | feature_matrix | scale + multilingual pretraining (full benchmarks pending technical report) | 40.4B params, decoder-only, 9.37T pretraining tokens, 35 European languages + 92 programming languages; mid-tier and below the 2026 frontier. The card reports 5-shot results for this model alone on BSC's own harness (English ARC-Challenge 59.4, ARC-Easy 86.5, PIQA 81.7) and defers comparisons with other models to a for... | null | null | low | ALIA-40B is a serious 40B multilingual base model, but it is smaller than Apertus-70B and below the 2026 frontier on reasoning and knowledge. It targets Iberian and European-language coverage rather than state-of-the-art capability. | null | 1.7 | adoption_and_capability | null | false | null | null | The record carries the current instruct build, ALIA-40b-instruct-2606, beside the base checkpoint. Openness is read on the base release. The earlier Salamandra family is not separately scored. | null | null | null | null | null | null | [
"https://github.com/langtech-bsc/alia"
] | null | null | null | null | [
"https://huggingface.co/BSC-LT/ALIA-40b",
"https://huggingface.co/BSC-LT/ALIA-40b-instruct-2606"
] | null | null | 15 |
marin | Marin | model | Marin is a fully open foundation-model line (Marin 8B Base and Instruct, Marin 32B Base) trained in JAX with the Levanter framework, with bit-for-bit reproducible training. Code, data mixtures, experiments, hyperparameters and training logs are published alongside the weights, and every experiment is tracked as a publi... | stanford-crfm | Stanford CRFM | null | base_pretrained | Base / pretrained models | model_components | Models | 26 | 5 | open_source | open | weights:open(Apache-2.0);data:open(documented mixtures);code:open(marin + Levanter, JAX);checkpoints:open;reproducibility:bit-for-bit;license:Apache-2.0(OSI) | high | Marin publishes Apache-2.0 weights together with its training code, data mixtures, experiments, hyperparameters and training logs, and its training runs are reproducible bit for bit. Nothing in the stack is withheld, so Marin is a fully open model. | 2026-11-12 | 1 | <10K | usage_volume | low | Adoption is measured as Hugging Face downloads summed across Marin's two declared artifacts, the 8B base and instruct checkpoints. It is a research-stage, fully-open project included on the map as an example of full openness rather than for its popularity. | 2026-11-08 | 3 | benchmark | MMLU/academic | Marin 32B Base averages 65.2 on its card's 19-task suite, ahead of OLMo 2 32B Base (63.2), level with Gemma 3 27B (65.1), behind Qwen 2.5 32B Base (68.1) | olmo | one_below | medium | On its own 19-task evaluation suite, Marin 32B Base scores about level with Gemma 3 27B and ahead of OLMo 2 32B, but trails Qwen 2.5 32B. The project has released no post-trained 32B or reasoning model to set against OLMo 3.1 Think, which keeps it a step below OLMo. | 2026-11-12 | 2.4 | adoption_and_capability | null | false | null | null | null | null | null | null | null | null | null | [
"https://github.com/marin-community/marin"
] | null | null | null | null | [
"https://huggingface.co/marin-community/marin-8b-base",
"https://huggingface.co/marin-community/marin-8b-instruct"
] | null | null | 13 |
granite | Granite | model | IBM's Granite family of enterprise language models, published under Apache-2.0. Granite 4.1, released in April 2026, added dense 3B, 8B and 30B models trained from scratch, with base and instruct checkpoints, and Granite 4.2 followed in August 2026, post-training those bases for reasoning, tool use and instruction foll... | ibm | IBM | null | base_pretrained | Base / pretrained models | model_components | Models | 27 | 3 | open_weights | open-ish | weights:open(Apache-2.0, ungated on HF across the 4.2 line);data:closed(the 4.1 pretraining mixture is described by category and phase and the 4.2 SFT sources by type, but no corpus is released);code:partial(the 4.2 and 4.1 language-model repositories carry documentation and usage examples, no pretraining pipeline);lic... | high | Granite 4.2, the current release, ships Apache-2.0 weights, ungated at every size, post-trained from the Granite 4.1 base models. IBM describes the 4.1 pretraining mixture by category and phase, and the 4.2 fine-tuning data by source type, but releases neither corpus, and its repositories carry documentation and usage ... | 2026-11-12 | 3 | 100K-1M | usage_volume | high | Adoption is measured as Hugging Face downloads summed across the declared Granite 4.0, 4.1 and 4.2 checkpoints, base and instruct. The small Granite 4.1 3B accounts for more than half of that reach. | null | 3 | benchmark | Artificial Analysis Intelligence Index v5.0.1 | Artificial Analysis Intelligence Index v5.0.1, read 2026-11-12: Granite 4.2 8B 5, eighth of ten small open-weight models (class median 8); Granite 4.2 3B 5, first of two tiny open-weight models (median 3). Granite 4.2 30B has no v5.0.1 score. Granite was last placed on Granite 4.0 h-small, level with Llama 4 Maverick o... | llama | at | medium | Granite 4.2, the current release, scores 5 at both 3B and 8B on version 5.0.1 of Artificial Analysis's Intelligence Index, the 3B leading the two tiny open-weight models it is compared with and the 8B near the bottom of its small-model class. Granite was last placed on Granite 4.0, level with Llama 4 Maverick on the ea... | null | 3 | adoption_and_capability | null | false | null | null | Openness follows Granite 4.2, the current release, which is post-trained from the Granite 4.1 base checkpoints. Adoption is read across the declared 4.0, 4.1 and 4.2 checkpoints. The speech, vision, guardian and switch models are separate lines and are not declared here. | null | null | null | null | null | null | [
"https://github.com/ibm-granite/granite-4.0-language-models",
"https://github.com/ibm-granite/granite-4.1-language-models",
"https://github.com/ibm-granite/granite-4.2-language-models"
] | null | null | null | null | [
"https://huggingface.co/ibm-granite/granite-4.0-h-small",
"https://huggingface.co/ibm-granite/granite-4.0-h-tiny",
"https://huggingface.co/ibm-granite/granite-4.0-h-micro",
"https://huggingface.co/ibm-granite/granite-4.0-micro",
"https://huggingface.co/ibm-granite/granite-4.1-3b",
"https://huggingface.co/... | null | null | 10 |
minicpm | MiniCPM | model | OpenBMB's on-device model family, spanning MiniCPM3-4B through the MiniCPM4/4.1 8B hybrid-reasoning generation to the current MiniCPM5-1B. The line pairs small dense models with efficiency research — trainable sparse attention (InfLLM-V2), quantized end-side inference — to run capable models on phones and edge hardware... | openbmb | OpenBMB | null | base_pretrained | Base / pretrained models | model_components | Models | 28 | 3 | open_weights | open-ish | weights:open(Apache-2.0, ungated on HF across generations 3-5);data:closed(training-data research published (Ultra-FineWeb, UltraClean) but the corpus is not released end to end);code:partial(OpenBMB/MiniCPM is inference/deployment plus papers; no full pretraining pipeline);license:Apache-2.0(OSI) | high | Apache-2.0 weights, ungated in every SKU read across generations 3 through 5, with training-data research published as papers and partial dataset artifacts but no end-to-end corpus release, and a repository that carries inference and deployment rather than the pretraining pipeline. Open weights, closed recipe: 3 by the... | null | 3 | 100K-1M | usage_volume | high | Adoption is measured as Hugging Face downloads summed across the four declared checkpoints spanning MiniCPM3 through the current MiniCPM5 generation. MiniCPM5-1B, the newest release, accounts for most of that total. | 2026-11-08 | 3 | benchmark | small-model suites (card comparisons vs similar-size Qwen3/Gemma) | MiniCPM4.1-8B is a hybrid reasoning model the card benchmarks against similar-size peers (Qwen3-8B class); the line's claim is best-in-class capability per parameter on device, not frontier capability. | smollm | at | medium | MiniCPM is competitive with same-size Qwen3 and Gemma checkpoints and is unusually efficient at long context via sparse attention, but the ceiling is the size class itself, well below the category's open-weight frontier. It is level with Smollm, the category's other small-model line, which makes the same claim: outperf... | 2026-11-12 | 3 | adoption_and_capability | null | false | null | null | null | null | null | null | null | null | null | [
"https://github.com/OpenBMB/MiniCPM"
] | null | null | null | null | [
"https://huggingface.co/openbmb/MiniCPM5-1B",
"https://huggingface.co/openbmb/MiniCPM4.1-8B",
"https://huggingface.co/openbmb/MiniCPM4-8B",
"https://huggingface.co/openbmb/MiniCPM3-4B"
] | null | null | 8 |
seed-oss | Seed-OSS | model | ByteDance Seed's open-weight model line, released August 2025 as 36B base and instruct checkpoints trained on 12T tokens. The instruct model targets long-context, reasoning and agentic use with a controllable thinking budget, and the base ships in with- and without-synthetic-data variants aimed at fine-tuning research. | bytedance-seed-volcano-engine | ByteDance Seed / Volcano Engine | null | base_pretrained | Base / pretrained models | model_components | Models | 29 | 3 | open_weights | open-ish | weights:open(36B Base, Base-woSyn and Instruct, ungated on HF);data:closed(12T-token pretraining corpus not released; the woSyn base variant is a synthetic-data ablation, not a data release);code:partial(ByteDance-Seed/seed-oss is inference/usage; no training pipeline);license:Apache-2.0(OSI) | high | Seed-OSS ships Apache-2.0 weights across all three SKUs, ungated, but the 12T-token training corpus is unreleased and only inference code is published. The model card states plainly that the series is released "under the Apache-2.0 license" while its technical report on the training recipe is still coming. | 2026-11-12 | 2 | 10K-100K | usage_volume | high | Adoption is measured as Hugging Face downloads summed across the two declared checkpoints, the instruct and base 36B releases. | 2026-11-08 | 4 | benchmark | AIME/GPQA/long-context suites (card benchmark tables) | Seed-OSS-36B-Instruct card benchmark tables: AIME24 91.7, AIME25 84.7-class math results and leading long-context/agentic scores among open-source peers in its size class, at 12T training tokens. | olmo | at | medium | The strong 30B-class open-weight tier: card benchmarks put the instruct model at or ahead of same-size open peers on math and reasoning despite a smaller training budget, which is the same "best of the ~32B open class, short of the trillion-parameter frontier" reading that puts olmo at 4. Vendor-published numbers with ... | 2026-11-12 | 3.4 | adoption_and_capability | null | false | null | null | ByteDance's open-weight counterpart to the closed doubao-seed line, and the openness contrast between the two is the org-page signal. The model card still marks the technical report as coming soon. | null | null | null | null | null | null | [
"https://github.com/ByteDance-Seed/seed-oss"
] | null | null | null | null | [
"https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct",
"https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Base"
] | null | null | 6 |
comma | Comma v0.1 | model | Comma v0.1 is a 7B base model trained on 2 trillion tokens drawn entirely from the Common Pile v0.1, an 8TB corpus of public domain and openly licensed text, and released with the paper that describes that corpus. It exists to test whether competitive models can be trained without unlicensed data, and performs comparab... | eleutherai | EleutherAI | null | base_pretrained | Base / pretrained models | model_components | Models | 30 | 5 | open_source | open | weights:open(Apache-2.0 safetensors, ungated (comma-v0.1-2t, -1t));data:open(comma_v0.1_training_dataset released ungated on HF; Common Pile v0.1 corpus public (8TB, openly licensed text));code:open(data pipeline in r-three/common-pile (MIT); training on facebookresearch/lingua (BSD-3-Clause) with the exact configs and... | high | Comma's Apache-2.0 weights ship alongside the exact training dataset, released ungated on the Hub, an MIT-licensed data pipeline, and a training run reproducible from a published config on the open BSD-3-Clause Lingua framework. Only intermediate checkpoints are missing, and the team notes it cannot guarantee the corpu... | 2026-11-12 | 1 | <10K | usage_volume | high | Adoption is measured as Hugging Face downloads summed across the two declared checkpoints. Comma is a research artifact rather than a production model, and its training dataset draws more downloads than the weights themselves. | 2026-11-08 | 2 | benchmark | pretraining suite (card table vs budget-matched peers) | Comma v0.1-2T 11-task average 57.4 (ARC/MMLU/BoolQ/HellaSwag/HumanEval/MBPP...), above OLMo Twin (51.6) and Llama 2 7B (55.8), below DeepSeekLLM 7B (58.8) — parity with budget-matched 2023-era 7B models trained on unlicensed data. | null | null | high | Comma reaches Llama 2-era parity, showing that openly licensed data alone can match budget-trained peers. That is a 2023-era capability bar, well below OLMo's current frontier and level with the category's other small fully-open research models, Lucie-7B and Luciole, ahead of Pythia. | 2026-11-12 | 1.7 | adoption_and_capability | null | false | null | null | null | null | null | null | null | null | null | [
"https://github.com/r-three/common-pile"
] | null | null | null | null | [
"https://huggingface.co/common-pile/comma-v0.1-2t",
"https://huggingface.co/common-pile/comma-v0.1-1t"
] | null | [
"https://arxiv.org/abs/2506.05209"
] | 10 |
eurollm | EuroLLM | model | EuroLLM is an open-weights multilingual LLM family from a European consortium led by Unbabel and DeepSPIN at Instituto Superior Técnico, under the UTTER Horizon Europe project. The 22B flagship, plus 9B and 1.7B siblings, is trained from scratch on ~4 trillion tokens across the EU's 24 official languages plus 11 more, ... | eurollm | EuroLLM | null | base_pretrained | Base / pretrained models | model_components | Models | 31 | 3 | open_weights | open-ish | weights:open(Apache-2.0, BF16 safetensors, ungated);data:documented-not-released(EuroWeb-2512 web subset released, remaining parallel/curated/synthetic and long-context components documented in the report but withheld);code:closed(only deep-spin/eurollm-eval evaluation harness is EuroLLM-specific, deep-spin/Megatron-LM... | high | EuroLLM ships ungated Apache-2.0 BF16 weights and publishes intermediate Megatron-Core checkpoints for every pretraining phase, matching what OLMo and Apertus release. Only the EuroWeb-2512 web subset of its training data is public -- the parallel, code, math, synthetic-math, book and long-context components are docume... | 2026-11-11 | 2 | 10K-100K | usage_volume | medium | Adoption is measured as trailing-30-day Hugging Face downloads summed across the ten declared EuroLLM SKUs -- base and instruct variants at 22B, 9B and 1.7B, current and preview -- on the utter-project org. Because downloads move from month to month across that many separate checkpoints, the reading carries only modera... | null | 3 | benchmark | MMLU / instruction-tuned English + multilingual suite | EuroLLM-22B-Instruct MMLU 69.8, MMLU-Pro 50.8, ARC-C 89.8, GSM8K 85.5 (English, Table 3); the strongest fully-open European model, ahead of Apertus on eight of ten English benchmarks (e.g. MMLU 69.8 vs Apertus-70B 67.9) at roughly one third of Apertus-70B's parameters, but trailing the frontier open-weights (Qwen-3-32B... | apertus | at | high | EuroLLM-22B beats Apertus, the strongest fully open European peer, on eight of ten English benchmarks in the technical report's own comparison, and its multilingual coverage across 35 languages is unusually strong. It still trails frontier open-weights models on reasoning and knowledge overall, so the win over Apertus ... | null | 2.7 | adoption_and_capability | null | false | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/utter-project/EuroLLM-22B-2512",
"https://huggingface.co/utter-project/EuroLLM-22B-Instruct-2512",
"https://huggingface.co/utter-project/EuroLLM-22B-Preview",
"https://huggingface.co/utter-project/EuroLLM-22B-Instruct-Preview",
"https://huggingface.co/utter-project/EuroLLM-9B",
"ht... | [
"https://huggingface.co/datasets/utter-project/EuroWeb-2512",
"https://huggingface.co/datasets/utter-project/EuroBlocks-SFT-2512"
] | [
"https://arxiv.org/abs/2602.05879"
] | 7 |
hunyuan | Hunyuan (Tencent Hy) | model | Hunyuan is Tencent's large language model line, released under the Tencent Hy name since 2026. Hy3 is a 295B-parameter mixture-of-experts model with 21B active parameters and a 256K-token context, and Hy4 preview scales the design to 770B total and 49B active parameters with a one-million-token context. Both ship as in... | tencent | Tencent | null | base_pretrained | Base / pretrained models | model_components | Models | 32 | 3 | open_weights | open-ish | weights:open(Hy4 preview and Hy3, with FP8 variants, ungated on Hugging Face);data:closed(no pretraining corpus or data listing is published);code:partial(fine-tuning and reinforcement-learning guides only, no pretraining pipeline);license:Apache-2.0(OSI);prior-release:distributed(Hunyuan-A13B, an earlier release, ship... | high | Hy4 preview and Hy3 ship their weights under Apache-2.0, ungated on Hugging Face, with fine-tuning and reinforcement-learning recipes alongside. The pretraining data and the pretraining pipeline are not published, which keeps the line at open weights. | 2026-11-12 | 4 | 1M-10M | usage_volume | high | Hugging Face downloads summed across the Hy3 and Hy4 preview checkpoints and their FP8 variants, nearly all of them for Hy3. | null | 4 | benchmark | vendor blind expert evaluations on the Hy3 and Hy4 preview model cards | Hy4 preview (770B total, 49B active): in Tencent's blind side-by-side rating of 203 engineering tasks by 163 internal experts it averaged 2.99, against 2.92 for GLM 5.3 and 2.94 for Kimi K3. Hy3: 2.67 out of 4 against 2.51 for GLM-5.1 in a 270-expert blind evaluation. Public benchmark tables appear only as images. | glm | one_below | low | Tencent's own blind evaluations put Hy4 preview level with GLM 5.3 and Kimi K3 on engineering work, but its public benchmark tables are published only as images, and the vendor calls the release an early version with known issues. Without an independent result it sits a step below GLM rather than beside it. | 2026-11-12 | 4 | adoption_and_capability | strong | false | null | null | Scored on the Hy releases. Tencent's video and 3D models are separate entries, hunyuan-video and hunyuan-3d. | null | null | null | null | null | null | [
"https://github.com/Tencent-Hunyuan/Hy4-preview",
"https://github.com/Tencent-Hunyuan/Hy3"
] | null | null | null | null | [
"https://huggingface.co/tencent/Hy4-preview",
"https://huggingface.co/tencent/Hy4-preview-FP8",
"https://huggingface.co/tencent/Hy3",
"https://huggingface.co/tencent/Hy3-FP8"
] | null | null | 13 |
snowflake-arctic | Snowflake Arctic | model | Snowflake Arctic is a 480B-parameter language model that pairs a 10B dense transformer with a residual mixture-of-experts layer of 128 experts, activating 17B parameters per token. Snowflake's AI research team trained it from scratch and released base and instruct checkpoints in April 2024, tuned for SQL generation, co... | snowflake | Snowflake | null | base_pretrained | Base / pretrained models | model_components | Models | 33 | 3 | open_weights | open-ish | weights:open(base and instruct checkpoints, ungated on Hugging Face);data:closed(the corpus is not released; the README points to a blog series on data rather than a dataset);code:partial(inference and vLLM tutorials; the repository's training code targets Llama 3.1 fine-tuning, not Arctic pretraining);license:Apache-2... | high | Arctic's base and instruct weights are published under Apache-2.0 without a gate. Snowflake describes its data approach in blog posts but releases neither the training corpus nor the pretraining code, so it stays at open weights. | 2026-11-12 | 1 | <10K | usage_volume | high | Hugging Face downloads summed across the base and instruct checkpoints. | null | 2 | benchmark | Snowflake's enterprise-metric comparison (HumanEval+, MBPP+, Spider, IFEval) in the release README | Arctic is on par with or better than Llama 3 8B and Llama 2 70B on Snowflake's enterprise metrics (HumanEval+, MBPP+, Spider, IFEval), on par with Llama 3 70B on those metrics, and competitive on language understanding and reasoning, per the release README. | internlm | at | medium | Snowflake reports Arctic matching Llama 3 70B on coding, SQL and instruction following, but on general understanding it claims only to stay competitive, which places it with the 2024 open models of the Llama 3 8B class where InternLM 2.5 also sits. It is well below current releases. | 2026-11-12 | 1.7 | adoption_and_capability | null | false | null | null | Arctic has had no new release since April 2024. Snowflake's later Arctic-named models are embedding, text-to-SQL and speculative-decoding checkpoints, and Arctic Embed is its own entry. | null | null | null | null | null | null | [
"https://github.com/Snowflake-Labs/snowflake-arctic"
] | null | null | null | null | [
"https://huggingface.co/Snowflake/snowflake-arctic-instruct",
"https://huggingface.co/Snowflake/snowflake-arctic-base"
] | null | null | 7 |
bert | BERT | model | BERT (Bidirectional Encoder Representations from Transformers) is Google's encoder-only language model, pretrained by masking words and predicting the next sentence on English Wikipedia and BookCorpus. Released from 2018 in Base (110M parameters) and Large (340M) sizes, cased and uncased, with later multilingual and Ch... | google-research | Google Research | null | base_pretrained | Base / pretrained models | model_components | Models | 34 | 4 | open_weights | open-ish | weights:open(Base, Large, multilingual and Chinese checkpoints, ungated on Hugging Face);data:documented-not-released(English Wikipedia and BookCorpus are named, but the preprocessed corpus was never released and BookCorpus is no longer distributed);code:open(create_pretraining_data.py and run_pretraining.py in google-... | high | BERT's checkpoints and its pretraining code, from data generation to the training loop, are published under Apache-2.0. The pretraining corpus is described but was never released, and BookCorpus is no longer offered by its collectors, so the original training run cannot be rebuilt exactly. | 2026-11-12 | 5 | >10M | usage_volume | high | Hugging Face downloads summed across Google's original BERT checkpoints, most of them for bert-base-uncased. | null | 1 | benchmark | GLUE and SQuAD results in the 2018 paper | GLUE score 80.5, MultiNLI accuracy 86.7, SQuAD v1.1 test F1 93.2 and SQuAD v2.0 test F1 83.1 at release (arXiv:1810.04805), the state of the art on eleven tasks in 2018. A 340M-parameter encoder that does not generate text. | comma | one_below | high | BERT set the state of the art on eleven language-understanding tasks in 2018, but it is an encoder of at most 340M parameters that cannot generate text, and the benchmarks used for today's language models do not apply to it. It sits a step below Comma, a 7B model that only matches 2023-era peers. | 2026-11-12 | 2.2 | adoption_and_capability | null | false | null | null | null | null | null | null | null | null | null | [
"https://github.com/google-research/bert"
] | null | null | null | null | [
"https://huggingface.co/google-bert/bert-base-uncased",
"https://huggingface.co/google-bert/bert-base-cased",
"https://huggingface.co/google-bert/bert-large-uncased",
"https://huggingface.co/google-bert/bert-large-cased",
"https://huggingface.co/google-bert/bert-base-multilingual-cased",
"https://huggingf... | null | [
"https://arxiv.org/abs/1810.04805"
] | 12 |
modernbert | ModernBERT | model | ModernBERT is an encoder-only language model from Answer.AI and LightOn that brings BERT's design up to date, with rotary position embeddings, alternating local and global attention and a native context of 8,192 tokens. It was pretrained on 2 trillion tokens of mostly English text and code and comes in Base (149M param... | answerdotai | Answer.AI | null | base_pretrained | Base / pretrained models | model_components | Models | 35 | 3 | open_weights | open-ish | weights:open(Base and Large checkpoints, ungated on Hugging Face);data:closed(2 trillion tokens of web documents, code and scientific literature, described only in those terms, with no sources or mixture named);code:open(pretraining and evaluation code in AnswerDotAI/ModernBERT);checkpoints:open(intermediate training c... | high | ModernBERT's weights, its training code and its intermediate checkpoints are all published under Apache-2.0. The training data is described only as web documents, code and scientific literature, without the sources or the mixture, so the corpus cannot be rebuilt. | 2026-11-12 | 4 | 1M-10M | usage_volume | high | Hugging Face downloads summed across the Base and Large checkpoints, most of them for Base. | null | 1 | benchmark | GLUE, BEIR, MLDR and code-retrieval results in the ModernBERT paper and model card | GLUE 88.4 (Base) and 90.4 (Large); BEIR 41.6 and 44.0 single-vector and 51.3 and 52.4 multi-vector; CodeSearchNet 56.4 and 59.5 (arXiv:2412.13663). A 395M-parameter encoder at most, which does not generate text. | comma | one_below | high | ModernBERT leads the older encoders it was tested against on most retrieval and code tasks, and its Base model is the first to pass DeBERTaV3 Base on GLUE, but it is an encoder of at most 395M parameters that cannot generate text. It sits a step below Comma, a 7B model that only matches 2023-era peers. | 2026-11-12 | 1.9 | adoption_and_capability | null | false | null | null | null | null | null | null | null | null | null | [
"https://github.com/AnswerDotAI/ModernBERT"
] | null | null | null | null | [
"https://huggingface.co/answerdotai/ModernBERT-base",
"https://huggingface.co/answerdotai/ModernBERT-large"
] | null | [
"https://arxiv.org/abs/2412.13663"
] | 13 |
roberta | RoBERTa | model | RoBERTa is Facebook AI's retraining of BERT: the same encoder-only architecture, trained longer, on larger batches and on about 160GB of English text, without BERT's next-sentence task and with masks that change on every pass over the data. Released in 2019 through the fairseq toolkit in Base (125M parameters) and Larg... | meta | Meta | null | base_pretrained | Base / pretrained models | model_components | Models | 36 | 4 | open_weights | open-ish | weights:open(Base and Large checkpoints, ungated on Hugging Face and on fairseq's download server);data:documented-not-released(BookCorpus, English Wikipedia, CC-News, OpenWebText and Stories, about 160GB, are named, but the assembled corpus was never released);code:open(fairseq ships the model and a recipe for pretrai... | high | RoBERTa's checkpoints and the fairseq code that trained them are published under the MIT license, with a written recipe for pretraining from scratch. The five corpora it trained on are named, but the 160GB mix was never released, so the original run cannot be rebuilt exactly. | 2026-11-12 | 5 | >10M | usage_volume | high | Hugging Face downloads summed across the Base and Large checkpoints, split fairly evenly between them. | null | 1 | benchmark | GLUE, SuperGLUE, SQuAD and RACE results in the 2019 paper and fairseq README | roberta.large: GLUE dev MNLI 90.2, SST-2 96.4 and CoLA 68.0, SQuAD 2.0 dev F1 89.4 and RACE test accuracy 83.2, state of the art on GLUE, RACE and SQuAD at release in 2019 (arXiv:1907.11692). A 355M-parameter encoder that does not generate text. | comma | one_below | high | RoBERTa showed that BERT had been undertrained and, retrained carefully, matched or beat every model published after it in 2019, but it is an encoder of at most 355M parameters that cannot generate text. It sits a step below Comma, a 7B model that only matches 2023-era peers. | 2026-11-12 | 2.2 | adoption_and_capability | null | false | null | null | XLM-RoBERTa, the multilingual model trained with the same recipe, was released separately and is not counted here. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/FacebookAI/roberta-base",
"https://huggingface.co/FacebookAI/roberta-large"
] | null | [
"https://arxiv.org/abs/1907.11692"
] | 10 |
t5 | T5 | model | T5 (Text-to-Text Transfer Transformer) is Google's encoder-decoder language model that treats every task, from translation and summarization to classification, as turning input text into output text. It was released in 2019 with the C4 web corpus it was pretrained on, in five sizes from Small (60M parameters) to 11B, a... | google-research | Google Research | null | base_pretrained | Base / pretrained models | model_components | Models | 37 | 5 | open_source | open | weights:open(Small, Base, Large, 3B and 11B checkpoints and the T5 1.1 set, ungated on Hugging Face);data:open(C4 is public and its TensorFlow Datasets builder recreates it from Common Crawl, and the supervised tasks mixed into the original checkpoints are public benchmarks defined in the t5 library);code:open(the t5 l... | high | T5's checkpoints and training code are published under Apache-2.0, and its C4 pretraining corpus is public, with a TensorFlow Datasets build that recreates it from Common Crawl. The supervised tasks mixed into the original checkpoints are public benchmarks, so the whole training run can be reproduced. | 2026-11-12 | 5 | >10M | usage_volume | high | Hugging Face downloads summed across the five original checkpoints and the five T5 1.1 checkpoints, most of them for T5-Small. | null | 1 | benchmark | summarization, question-answering and classification results in the 2019 paper | State of the art on many benchmarks covering summarization, question answering and text classification at release in 2019 (arXiv:1910.10683); the largest checkpoint has 11B parameters. | comma | one_below | medium | T5 set the state of the art on many language benchmarks in 2019 and, unlike an encoder, it generates text, but it has no results on the benchmarks used for today's language models and is mostly run in its smaller sizes. It sits a step below Comma, a 7B model that only matches 2023-era peers. | 2026-11-12 | 2.2 | adoption_and_capability | null | false | null | null | mT5, ByT5 and Flan-T5 were released separately as their own models and are not counted here. | null | null | null | null | null | null | [
"https://github.com/google-research/text-to-text-transfer-transformer"
] | null | null | null | null | [
"https://huggingface.co/google-t5/t5-small",
"https://huggingface.co/google-t5/t5-base",
"https://huggingface.co/google-t5/t5-large",
"https://huggingface.co/google-t5/t5-3b",
"https://huggingface.co/google-t5/t5-11b",
"https://huggingface.co/google/t5-v1_1-small",
"https://huggingface.co/google/t5-v1_1... | null | [
"https://arxiv.org/abs/1910.10683"
] | 19 |
deberta | DeBERTa | model | DeBERTa is Microsoft's encoder-only language model, which keeps each word's content and position apart inside attention. Its third version, DeBERTaV3, replaces masked-word prediction with ELECTRA-style replaced-token detection and comes in XSmall to Large sizes, plus mDeBERTaV3 Base, a multilingual model covering 102 l... | microsoft | Microsoft | null | base_pretrained | Base / pretrained models | model_components | Models | 38 | 4 | open_weights | open-ish | weights:open(DeBERTaV3 XSmall, Small, Base and Large and mDeBERTaV3 Base, ungated on Hugging Face);data:documented-not-released(the English models use the 160GB mix of Wikipedia, BookCorpus, OpenWebText, Stories and CC-News shared with DeBERTa and RoBERTa, named but never released, while mDeBERTaV3 uses the public CC10... | high | DeBERTaV3's checkpoints and its pretraining scripts are published under the MIT license. The English models were trained on the same 160GB mix of encyclopedia, book, news and web text as RoBERTa, which is described in the papers but was never released as a corpus. | 2026-11-12 | 4 | 1M-10M | usage_volume | high | Hugging Face downloads summed across Microsoft's DeBERTa checkpoints, most of them for mDeBERTaV3 Base and the English DeBERTaV3 Base. | null | 1 | benchmark | GLUE and XNLI results in the DeBERTaV3 paper | GLUE average 91.37 for DeBERTaV3 Large and 79.8 zero-shot cross-lingual XNLI accuracy for mDeBERTaV3 Base, the state of the art among models of similar structure at release (arXiv:2111.09543). A 304M-backbone encoder at most, which does not generate text. | comma | one_below | high | DeBERTaV3 Large posted the best GLUE result of any encoder its size when it was released, and the multilingual Base model led XNLI, but these are encoders that cannot generate text. It sits a step below Comma, a 7B model that only matches 2023-era peers. | 2026-11-12 | 1.9 | adoption_and_capability | null | false | null | null | null | null | null | null | null | null | null | [
"https://github.com/microsoft/DeBERTa"
] | null | null | null | null | [
"https://huggingface.co/microsoft/deberta-v3-base",
"https://huggingface.co/microsoft/deberta-v3-large",
"https://huggingface.co/microsoft/deberta-v3-small",
"https://huggingface.co/microsoft/deberta-v3-xsmall",
"https://huggingface.co/microsoft/mdeberta-v3-base",
"https://huggingface.co/microsoft/deberta... | null | [
"https://arxiv.org/abs/2111.09543",
"https://arxiv.org/abs/2006.03654"
] | 21 |
ernie-flash | ERNIE Flash | model | ERNIE Flash is Baidu's line of smaller, efficient ERNIE models. Its open release, ERNIE 5.5 Flash Base, published on Hugging Face in October 2026, is a mixture-of-experts base model with 280B total and 14B active parameters, pretrained from scratch and released at the end of pretraining so developers can continue train... | baidu | Baidu | null | base_pretrained | Base / pretrained models | model_components | Models | 39 | 3 | open_weights | open-ish | weights:open(FP8 and BF16 safetensors, ungated on Hugging Face);data:closed(the pretraining data is described by category (web, code, long context), with no corpus, listing or mixture released);code:partial(inference through SGLang and the vLLM-ERNIE plugin; no pretraining pipeline);license:ERNIE-5.5-Flash-Base-Model-L... | high | Baidu publishes the ERNIE 5.5 Flash Base weights on Hugging Face, ungated, under its own license, which allows commercial use without a fee up to US$50 million in related annual revenue and requires written permission above that. The license does not cover the EU, the UK or South Korea. Baidu describes its pretraining ... | 2026-11-12 | 1 | <10K | usage_volume | high | Hugging Face downloads summed across the FP8 and BF16 copies of ERNIE 5.5 Flash Base, released in late October 2026. | null | 4 | benchmark | Baidu's same-pipeline pretraining-suite table on the ERNIE 5.5 Flash Base card | On Baidu's own evaluation of base checkpoints under one pipeline, ERNIE 5.5 Flash Base (280B total, 14B active) leads Hy3-preview-Base (295B, 21B active) on 27 of 30 benchmarks, including MMLU-Pro 78.4 vs 66.9, GPQA 60.2 vs 50.2 and LiveCodeBench 40.1 vs 33.5, and leads DeepSeek-V4-Flash-Base and MiMo-V2.5-Base on most... | hunyuan | at | low | Baidu's own comparison of base checkpoints, run under a single evaluation pipeline, puts ERNIE 5.5 Flash Base ahead of the base model behind Tencent's Hy3 on almost every row, and ahead of DeepSeek's and Xiaomi's similar-size bases on most. The results are vendor-run and cover base models only, with no independent lead... | 2026-11-12 | 3.1 | adoption_and_capability | null | false | null | null | Scored on ERNIE 5.5 Flash Base, the open pretraining checkpoint. The API-only ERNIE 5.5 Flash Preview is left out of the openness reading, and Baidu's flagship line is the separate ernie entry. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/baidu/ERNIE-5.5-Flash-Base",
"https://huggingface.co/baidu/ERNIE-5.5-Flash-Base-BF16"
] | null | null | 8 |
qwen-coder | Qwen3-Coder | model | Qwen3-Coder is a line of open-weight models for agentic coding, tool calling and repository-scale code work. It spans Qwen3-Coder-Next, an 80B mixture-of-experts model with 3B active parameters built for coding agents and local development, and the larger Qwen3-Coder-480B-A35B-Instruct, with 35B of its 480B parameters ... | alibaba-cloud | Alibaba Cloud | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 0 | 3 | open_weights | open-ish | weights:open(Apache-2.0, on HF);data:closed(no training corpus released);code:partial(inference/usage + chat template; no training pipeline or post-training/SFT data);license:Apache-2.0(OSI permissive, no use restrictions) | high | Qwen3-Coder-Next, the newest release in the line, ships Apache-2.0 weights with no use restrictions, as the earlier Qwen3-Coder and Qwen2.5-Coder checkpoints do. The training corpus is closed, and only inference code and a chat template are released, with no training pipeline. That keeps the line at open weights rather... | 2026-11-08 | 4 | 1M-10M | usage_volume | high | This reading is carried over from the last scheduled measurement, which summed Hugging Face downloads across the four earlier Qwen2.5-Coder and Qwen3-Coder checkpoints. Qwen3-Coder-Next and its base model have since been added to the declared checkpoints, and a measurement of the expanded set is pending. Qwen's FP8 and... | null | 4 | benchmark | SWE-bench/agentic-coding | Vendor positions agentic coding comparable to Claude Sonnet class; 480B-A35B MoE, repository-scale long context; non-thinking mode | null | null | medium | Qwen's model card says the model achieves results comparable to Claude Sonnet on agentic coding among open models; the claim is vendor-reported and no independent SWE-bench result is cited. It sits a step below current frontier coding models. | null | 4 | adoption_and_capability | strong | false | null | [
"qwen2-5-coder-instruct",
"qwen3-coder"
] | Scored as one line across the Qwen2.5-Coder, Qwen3-Coder and Qwen3-Coder-Next releases. The Qwen3-Coder-Next model card gives no release date; its repository was created on 2026-01-30. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct",
"https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct",
"https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct",
"https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8",
"https://huggingface.co/Qwen/Qwen3-Coder-Next",
"https://huggingf... | null | null | 7 |
qwen-instruct | Qwen 3.8 | model | Alibaba's post-trained Qwen chat models. The current release, Qwen 3.8 from August 2026, ships three open-weight models that think by default and can answer directly on request: a dense 27B under Apache-2.0, Qwen3.8-Flash-Next, a 125B mixture of experts with 6B active parameters, and the 2.4-trillion-parameter Qwen3.8-... | alibaba-cloud | Alibaba Cloud | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 1 | 3 | open_weights | open-ish | weights:open(ungated on the Hub: Qwen3.8-27B, Qwen3.8-Flash-Next and the Qwen3.8-2.4T-A95B flagship);data:closed(no post-training data published; the cards name no datasets, and the Flash-Next technical report covers architecture, optimization and evaluation);code:partial(usage and deployment instructions in QwenLM/Qwe... | high | Qwen 3.8 ships its dense 27B chat model under Apache-2.0 and its Flash-Next and flagship models under Qwen's own licenses, which require model-as-a-service and AI coding or office assistant businesses to get a separate license before commercial use. Both licenses also require a product with more than 100 million monthl... | 2026-10-09 | 5 | >10M | usage_volume | high | Adoption is measured as Hugging Face downloads summed across the eight declared chat checkpoints, from Qwen2.5-14B-Instruct through the three Qwen 3.8 models. Qwen3.8-27B alone accounts for about two fifths of the total. | 2026-11-08 | 4 | benchmark | Artificial Analysis Intelligence Index | Qwen3.8-2.4T-A95B, the open-weight flagship of Qwen 3.8, scores 40 on the Artificial Analysis Intelligence Index against a median of 18 for open-weights models of its size; Qwen3.8-Flash-Next also scores 40 and the dense Qwen3.8-27B (Xhigh) 34. Gemini 3.8 Flash (High) scores 41. | gemini-flash | at | high | This score reads Qwen 3.8, the current release of Qwen's chat models. On the Artificial Analysis index its flagship sits a point behind Gemini 3.8 Flash and well behind the closed leaders, and even the dense 27B model places above Gemini 3.1 Pro, so it scores level with Gemini Flash. | 2026-10-09 | 4.3 | adoption_and_capability | strong | false | null | [
"qwen2-5-14b-instruct",
"qwen-3-6",
"qwen-3-7-max",
"qwen-3-8"
] | Covers Qwen's post-trained chat models; the base checkpoints are in qwen, the same split as llama-instruct and llama. Qwen2.5-14B-Instruct, the release this record was first written for, stays declared. | null | null | null | null | null | null | [
"https://github.com/QwenLM/Qwen3.8"
] | null | null | null | null | [
"https://huggingface.co/Qwen/Qwen3.8-27B",
"https://huggingface.co/Qwen/Qwen3.8-Flash-Next",
"https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B",
"https://huggingface.co/Qwen/Qwen3.6-27B",
"https://huggingface.co/Qwen/Qwen3.6-35B-A3B",
"https://huggingface.co/Qwen/Qwen3.5-397B-A17B",
"https://huggingface.co... | null | null | 21 |
deepseek-r1 | DeepSeek-R1 | model | DeepSeek-R1 is a reasoning model, released in January 2025, that uses large-scale GRPO reinforcement learning to develop chain-of-thought before answering. It is a 671B-parameter mixture-of-experts model with 37B active parameters and a 128K context window, post-trained on DeepSeek-V3-Base and released alongside 1.5B t... | deepseek | DeepSeek | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 2 | 3 | open_weights | open-ish | weights:open(MIT, on HF; distillation explicitly permitted);data:closed(RL/SFT training data not released);code:partial(inference; no full RL training pipeline);license:MIT(OSI-style permissive, no use restrictions) | high | MIT-licensed open weights that explicitly permit distillation, but the RL and SFT post-training data and the full training code are not released, so this is open weights rather than open source. | 2026-09-29 | 4 | 1M-10M | usage_volume | high | Adoption is measured as Hugging Face downloads on the base R1 repository, summed with the declared R1-0528 SKU. That count leaves out the large family of R1 distills and the DeepSeek app and API surface, so it captures only part of the release's total use, which drove widespread open-reasoning adoption. | 2026-11-08 | 4 | benchmark | AIME/MATH-500/GPQA/SWE-bench | AIME 2024 79.8, MATH-500 97.3, GPQA Diamond 71.5, MMLU 90.8, SWE-bench Verified 49.2, LiveCodeBench 65.9 (HF card, vs OpenAI o1) | null | null | high | DeepSeek-R1 was a frontier open reasoning model at its January 2025 release, matching or beating OpenAI's o1 on AIME, MATH-500 and LiveCodeBench; the figures are vendor-reported on its Hugging Face card. Newer DeepSeek releases, including the V4 family, have since surpassed it. | 2026-09-29 | 4 | adoption_and_capability | strong | false | null | null | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/deepseek-ai/DeepSeek-R1",
"https://huggingface.co/deepseek-ai/DeepSeek-R1-0528"
] | null | null | 5 |
deepseek-coder | DeepSeek-Coder-V2-Instruct | model | Dedicated coding mixture-of-experts model, instruction-tuned for generation, completion and chat, built on a DeepSeek-V2 checkpoint with six trillion additional code tokens. It ships in a 236B-total variant with 21B active parameters and a 16B Lite variant, supports 338 programming languages and a 128K context. | deepseek | DeepSeek | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 3 | 3 | open_weights | open-ish | weights:open(downloadable on HF);data:closed;code:open(MIT, repo code);model-license:DeepSeek-Model-License(non-OSI; Attachment A acceptable-use restrictions - military/illegal/discrimination/etc.; commercial allowed but use-based restrictions must propagate to derivatives) | high | DeepSeek releases the inference code under MIT, and the weights download without a gate, but the DeepSeek Model License's Attachment A bars military, illegal and discriminatory use and binds derivatives to the same restriction. That limits conduct rather than commerce, keeping the model at open weights rather than rest... | 2026-09-29 | 3 | 100K-1M | usage_volume | high | Adoption is measured as Hugging Face downloads summed across the two declared DeepSeek-Coder-V2 checkpoints, and nearly all of that volume goes to the smaller Lite variant rather than the 236B flagship. Later DeepSeek releases and 2026-generation coding models have drawn usage away from this line, and the flagship chec... | 2026-10-25 | 3 | benchmark | code-tasks | Vendor: comparable to GPT-4 Turbo on code-specific tasks, beats Claude 3 Opus / Gemini 1.5 Pro on coding+math at 2024 release; 338-language support | null | null | medium | DeepSeek's model card claims the model performs comparably to GPT-4 Turbo on code-specific tasks, and defers exact HumanEval, MBPP and SWE-bench numbers to the paper. It is a 2024-generation model and sits below current frontier coding models. | null | 3 | adoption_and_capability | null | false | null | [
"deepseek-coder-v2-instruct"
] | A 2024-generation coder, superseded by the V3 and V4 lines. | null | null | null | null | null | null | [
"https://github.com/deepseek-ai/DeepSeek-Coder-V2"
] | null | null | null | null | [
"https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct",
"https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Instruct"
] | null | null | 7 |
deepseek-instruct | DeepSeek-V4-Pro / V4-Flash | model | DeepSeek-V4-Pro and V4-Flash are open-weight mixture-of-experts language models with a one-million-token context window. V4-Pro has 1.6T total parameters (49B active) and V4-Flash has 284B (13B active). Both are published on Hugging Face by DeepSeek. | deepseek | DeepSeek | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 4 | 3 | open_weights | open-ish | weights:open(MIT, on HF);data:closed;code:partial(inference/serving; no training pipeline or data);license:MIT(OSI-style permissive, no use restrictions) | high | V4-Flash ships MIT-licensed weights, matching the V4-Pro flagship, but the training data stays closed and only inference and serving code is public. That keeps it an open-weights release rather than open source, the same class as V4-Pro. | 2026-09-29 | 4 | 1M-10M | usage_volume | high | Adoption sums Hugging Face downloads across the two declared checkpoints, V4-Pro and V4-Flash, as the family rule requires; V4-Flash draws roughly twice V4-Pro's downloads. | 2026-11-08 | 4 | benchmark | MMLU-Pro/LiveCodeBench | Instruct (Max mode): MMLU-Pro 86.2, LiveCodeBench 91.6, MRCR-1M 78.7; base MMLU 88.7, HumanEval 69.5, GSM8K 90.8 (HF card) | null | null | medium | V4-Flash scores near the frontier on DeepSeek's own benchmarks (MMLU-Pro 86.2, LiveCodeBench 91.6), figures reported on its Hugging Face model card. It is deliberately the lighter, faster sibling of the higher-scoring V4-Pro flagship, so Flash trails Pro rather than matching it. | 2026-09-29 | 4 | adoption_and_capability | strong | false | null | [
"deepseek-v4-pro-v4-flash"
] | Declares both instruct checkpoints, V4-Pro and V4-Flash, and sums their downloads for adoption. The deepseek record in base_pretrained declares the same two alongside the V3 and base checkpoints. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro",
"https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
] | null | null | 5 |
codellama | CodeLlama Instruct | model | Code Llama Instruct is a code-specialized, instruction-tuned family built on Llama 2 at 7B, 13B, 34B and 70B and fine-tuned for coding chat. Meta trained it during 2023. | meta | Meta | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 5 | 3 | open_weights | open-ish | weights:open(downloadable on HF);data:closed;code:partial(inference utils, no training pipeline);license:Llama-2-Community(custom Meta license, non-OSI; AUP + commercial conditions) | high | Code Llama's weights are downloadable under Meta's Llama 2 Community License, which is not OSI-approved. The license bounds rather than forbids commercial use, since anyone within its limits may use the model commercially, unlike a license such as CC-BY-NC or Mistral's Non-Production License that forbids commerce outri... | 2026-09-29 | 2 | 10K-100K | usage_volume | medium | Adoption is measured as Hugging Face downloads summed across the four Code Llama Instruct checkpoints this tier covers, most of them for the 7B model. All four repositories are declared on the product, and downloads are combined across the sizes rather than read from one checkpoint alone. | 2026-11-08 | 3 | benchmark | HumanEval | Code Llama - Instruct 13B HumanEval pass@1 42.7%, MBPP pass@1 49.4%; 70B-Instruct higher at 67.8% HumanEval, 34B-Instruct lower at 41.5%; all 2023-era | null | null | medium | Code Llama Instruct is a 2023-era code model that sits well below current coding models. Its Instruct sizes are not monotonic: the paper gives 13B 42.7% HumanEval / 49.4% MBPP, with 70B higher at 67.8% and 34B lower at 41.5%. | null | 2.7 | adoption_and_capability | null | false | null | [
"codellama-instruct"
] | Scored on the 13B Instruct release as representative. Trained during 2023 and superseded by later coder families; the model card inlines no benchmark numbers, so capability is read from the paper. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/codellama/CodeLlama-34b-Instruct-hf",
"https://huggingface.co/codellama/CodeLlama-7b-Instruct-hf",
"https://huggingface.co/codellama/CodeLlama-13b-Instruct-hf",
"https://huggingface.co/codellama/CodeLlama-70b-Instruct-hf"
] | null | null | 10 |
llama-instruct | Llama-3.1-8B-Instruct | model | Llama-3.1-8B-Instruct is an instruction-tuned 8-billion-parameter Llama 3.1 model for chat and assistant workflows, post-trained with supervised fine-tuning and reinforcement learning with human feedback on Meta's Llama 3.1 base. It supports a 128K-token context window. Weights are gated behind a manual access request ... | meta | Meta | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 6 | 3 | open_weights | open-ish | weights:open(downloadable);data:closed;code:partial(no training pipeline);license:Llama-3.1-Community(non-OSI; 700M-MAU commercial cap + use policy) | high | Meta releases the weights under the Llama 3.1 Community License, non-OSI and requiring a separate license from Meta above 700 million monthly active users. That cap permits commercial use below the threshold rather than forbidding it, so the release counts as open weights rather than fully open source: the training dat... | 2026-09-29 | 4 | 1M-10M | usage_volume | high | Adoption is read from Hugging Face downloads on the sole declared checkpoint, meta-llama/Llama-3.1-8B-Instruct. No other artifact is declared for this record, so the reading does not draw on any additional distribution surface that might place it higher. | 2026-11-08 | 3 | benchmark | MMLU/IFEval | Llama 3.1 8B-Instruct ~MMLU ~69, solid 8B chat; no reasoning/agentic frontier scores | null | null | medium | A competent 2024-generation 8B chat model, a long way below the 2026 frontier and below the larger Llama 3.1 SKUs. The card's benchmark table gives Llama 3.1 8B Instruct MMLU 69.4 (5-shot, macro_avg/acc) against 83.6 for the 70B and 87.3 for the 405B, and publishes no reasoning or agentic frontier scores. | 2026-09-29 | 3.3 | adoption_and_capability | null | false | null | [
"llama-3-1-8b-instruct"
] | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct"
] | null | null | 5 |
devstral | Devstral | model | Devstral is an open-weight family of agentic coding models from Mistral, built for software-engineering tasks such as exploring codebases and editing multiple files. Devstral Small 2 is a 24B instruct model with a 256K context window that runs on a single RTX 4090 or a Mac with 32GB RAM, and the 123B Devstral 2 is the ... | mistral-ai | Mistral AI | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 7 | 3 | open_weights | open-ish | weights:open(Apache-2.0, on HF);data:closed;code:partial(inference/serving; no training or post-training data);license:Mistral-Modified-MIT(non-OSI; withholds every right from a company whose consolidated monthly revenue exceeds $20 million; resolved most-restrictive across the governing release's SKUs (Devstral 2 123B... | high | Devstral 2's smaller 24B model ships under Apache-2.0, but the larger 123B model carries a non-OSI Modified MIT license; since the two are sizes of one release, the record follows the more restrictive license. Training and post-training data and the pipeline behind the checkpoints are undisclosed, which keeps the famil... | 2026-09-29 | 3 | 100K-1M | usage_volume | high | Devstral is a specialist coding-agent model, adopted mainly through OpenHands and similar coding-agent tooling rather than general chat use. Adoption is measured as Hugging Face downloads summed across the two declared Devstral 2 checkpoints, most of it on the smaller 24B model rather than the 123B flagship. | 2026-11-08 | 4 | benchmark | SWE-bench Verified | Devstral 2 123B SWE-bench Verified 72.2% / SWE-bench Multilingual 61.3% / Terminal-Bench 2 32.6%; Devstral Small 2 24B 68.0% / 55.7% / 22.5% | null | null | high | Devstral 2's 123B model scores 72.2% on SWE-bench Verified on its model card, ahead of Qwen3 Coder Plus, Kimi K2 Thinking and GLM 4.6, while the smaller 24B model scores 68.0%. DeepSeek v3.2 and the proprietary models on the same table score higher, and Mistral's newer Medium 3.5 model has replaced Devstral 2 as its co... | 2026-09-29 | 3.7 | adoption_and_capability | null | false | null | null | Scored against the Devstral 2 release; the earlier Small 1.1 checkpoints are superseded. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct-2512",
"https://huggingface.co/mistralai/Devstral-2-123B-Instruct-2512"
] | null | null | 9 |
mistral-7b-instruct | Mistral-7B-Instruct-v0.2 | model | Mistral-7B-Instruct-v0.2 is an instruction-tuned 7B chat model with a 32K context window, fine-tuned from the Mistral-7B-v0.2 base model and prompted with [INST] tags. Mistral describes it as a quick demonstration that the base model can be fine-tuned easily, and the card notes it has no moderation mechanisms. | mistral-ai | Mistral AI | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 8 | 3 | open_weights | open-ish | weights:open(Apache-2.0, on HF);data:closed;code:closed(no training/post-training pipeline);license:Apache-2.0(OSI permissive) | high | Apache-2.0 open weights, with the instruction-tuning data and recipe undisclosed, so this is open weights rather than open source. | 2026-09-29 | 4 | 1M-10M | usage_volume | high | Adoption is read from Hugging Face downloads on the single declared SKU, Mistral-7B-Instruct-v0.2, which also carries a large derivative ecosystem of finetunes, adapters and merges. The checkpoint remains heavily used as a permissive 7B instruct base years after its release. | 2026-11-08 | 2 | benchmark | MT-Bench/MMLU | Mistral 7B Instruct v0.2 MT-Bench 7.6 (Mistral-reported); underlying Mistral 7B base MMLU 60.1%; v0.1-Instruct MT-Bench 6.84. 2023-era 7B chat quality | null | null | medium | Mistral reports 7.6 on MT-Bench for v0.2, and the Mistral 7B paper gives 6.84 for the original v0.1 instruct model, which it says outperformed all 7B models on that benchmark. Later Mistral 7B releases have replaced this version. | 2026-09-29 | 2.6 | adoption_and_capability | null | false | null | [
"mistral-7b-instruct-v0-2"
] | Superseded by Mistral-7B-Instruct-v0.3. The card carries no benchmark table and links only the base-model paper, so the capability figures come from the paper and the La Plateforme launch post. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2"
] | null | null | 6 |
starcoder2 | StarCoder2 | model | BigCode's self-aligned instruct release of StarCoder2 at 16B, trained with a fully transparent pipeline and evaluated on EvalPlus, LiveCodeBench and DS-1000. The base StarCoder2-15B it builds on was trained on The Stack v2 and belongs to the base_pretrained shelf. | bigcode-project | BigCode Project | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 9 | 4 | open_weights | open-ish | weights:open;data:open(self-oss-instruct-sc2-exec-filter-50k + 7 documented intermediate datasets);code:open(starcoder2-self-align pipeline on GitHub);base-data:open(The Stack v2);license:BigCode-OpenRAIL-M(open but RAIL behavioral-use restrictions, not OSI) | high | StarCoder2 publishes its weights, its full training corpus (The Stack v2) and the self-alignment pipeline behind the instruct checkpoint. Its BigCode-OpenRAIL-M license restricts what the model can be used for rather than who can use it commercially, and because that license is not OSI-approved, it stays at open weight... | 2026-09-29 | 1 | <10K | usage_volume | medium | Adoption is measured as Hugging Face downloads for the instruct SKU and the base StarCoder2-15B, and both see minimal download volume. The self-alignment recipe is significant to researchers, but that has not translated into wide production use. | 2026-09-27 | 3 | benchmark | HumanEval | StarCoder2-15B-Instruct-v0.1 HumanEval pass@1 72.6% (HumanEval+ 63.4%) | null | null | high | The model card reports HumanEval pass@1 of 72.6 and HumanEval+ of 63.4 and calls it the first entirely self-aligned code model trained with a fully permissive and transparent pipeline. It reports no agentic coding results, and it sits below current frontier coding models. | 2026-09-29 | 2.4 | adoption_and_capability | null | false | null | null | Scored on the instruct model; the declared artifact is the base repository, so its downloads are read as the instruct model's. | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/bigcode/starcoder2-15b"
] | null | null | 6 |
olmo-instruct | OLMo 3 Instruct | model | Instruction-tuned chat releases of Ai2's OLMo 3 family at 7B and 32B, post-trained on the fully open OLMo 3 base with the open Tulu and OLMo recipe. Every stage, checkpoint, dataset and dependency needed to recreate or modify them is published. It supersedes OLMo 2 Instruct. | allen-institute-for-ai | Allen Institute for AI | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 10 | 5 | open_source | open | weights:open(Apache-2.0);data:open(Dolma 3 pretraining + open post-training mixtures);code:open(OLMo + open-instruct);checkpoints:open;license:Apache-2.0(OSI) | high | A fully open instruct line: Apache-2.0 weights, open pretraining and post-training data, open training code and open checkpoints. The line now extends to an Olmo 3.1 update shipping 32B Think and Instruct checkpoints on the same open recipe, which does not change the class. | 2026-09-29 | 3 | 100K-1M | usage_volume | medium | Adoption is measured as Hugging Face downloads summed across the two declared releases, Olmo-3-7B-Instruct and Olmo-3.1-32B-Instruct. The 32B checkpoint declared before it, Olmo-3-1025-32B-Instruct, is no longer on the Hub, and 3.1 is the 32B Instruct Ai2 publishes now. | 2026-11-08 | 4 | benchmark | IFEval/AlpacaEval/MMLU | OLMo 3-Instruct 7B ties or surpasses Qwen 2.5, Gemma 3 and Llama 3.1 at its scale | null | null | high | Ai2's evaluations show Olmo 3 Instruct tying or surpassing Qwen 2.5, Gemma 3 and Llama 3.1 at similar scale and competing with the Qwen 3 family, and the 3.1 32B Instruct is described by Ai2 as the strongest fully open 32B-scale instruct model. It is a fully open line that stays below the current frontier of proprietar... | 2026-09-29 | 3.7 | adoption_and_capability | null | false | null | [
"olmo-3-instruct"
] | Adoption counts the two declared instruct checkpoints, 7B and 3.1 32B, and not their SFT, DPO or quantized variants, so it is a floor. | null | null | null | null | null | null | [
"https://github.com/allenai/OLMo",
"https://github.com/allenai/open-instruct"
] | null | null | null | null | [
"https://huggingface.co/allenai/Olmo-3-7B-Instruct",
"https://huggingface.co/allenai/Olmo-3.1-32B-Instruct"
] | null | null | 5 |
olmoe-instruct | OLMoE Instruct | model | Ai2's fully open mixture-of-experts instruct model, with roughly 1 billion active parameters out of 6.9 billion total, post-trained with supervised fine-tuning and DPO. It was released with its pretraining data, post-training data, training code, logs and 244 intermediate checkpoints. | allen-institute-for-ai | Allen Institute for AI | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 11 | 5 | open_source | open | weights:open(Apache-2.0);data:open(OLMoE-mix-0924 pretraining + Tulu/UltraFeedback post-training);code:open(OLMoE training code, SFT/DPO/KTO recipes);checkpoints:open(244 intermediate);license:Apache-2.0(OSI) | high | A fully open mixture-of-experts model: Apache-2.0 weights, open pretraining data, open post-training data, open training code and 244 intermediate checkpoints. | 2026-09-29 | 2 | 10K-100K | usage_volume | high | Adoption is measured as Hugging Face downloads summed across the two declared OLMoE artifacts, the instruct model and its pretraining mixture dataset. Nearly all of that volume goes to the model itself rather than to the dataset download. | 2026-10-25 | 2 | benchmark | MMLU/AlpacaEval | OLMoE-1B-7B-Instruct MMLU 51.9, GSM8k 45.5, HumanEval 54.8, AlpacaEval1 84.0, IFEval 48.1 | null | null | medium | The model card calls it state-of-the-art among models of similar cost, about 1B active parameters, and its own comparison table puts it ahead of OLMo-7B, DeepSeek-3B-16B and Qwen1.5-3B-14B chat models on average score. Its small active-parameter budget keeps it well below full-size current chat models on absolute capab... | 2026-09-29 | 2 | adoption_and_capability | null | false | null | null | null | null | null | null | null | null | null | [
"https://github.com/allenai/OLMoE"
] | null | null | null | null | [
"https://huggingface.co/allenai/OLMoE-1B-7B-0924-Instruct"
] | [
"https://huggingface.co/datasets/allenai/OLMoE-mix-0924"
] | null | 12 |
tulu | Tulu 3 | model | Ai2's open post-training model family at 8B, 70B and 405B, applying a fully reproducible recipe on top of Meta's Llama 3.1 base: open supervised and preference data, the open-instruct training code, and reinforcement learning with verifiable rewards. The evaluation framework is released alongside it. | allen-institute-for-ai | Allen Institute for AI | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 12 | 3 | open_weights | open-ish | weights:open(downloadable);license:Llama-3.1-Community(non-OSI;700M-MAU cap + use policy);post-training-data:open(Apache-2.0/ODC-BY);code:open(open-instruct Apache-2.0);recipe:open(SFT->DPO->RLVR);evals:open(OLMES) | high | The released weights inherit Meta's Llama 3.1 Community License, non-OSI and capped at 700 million monthly active users for commercial use, governing the checkpoint regardless of what Ai2 adds. The post-training layer above it is fully open under Apache-2.0: the data, the open-instruct code, the SFT-DPO-RLVR recipe and... | 2026-09-29 | 1 | <10K | usage_volume | high | Adoption sums Hugging Face downloads across Tulu's three declared model checkpoints, 8B, 70B and 405B, most of the volume on the 8B. The tulu-3-sft-mixture dataset the product also declares is scored separately as its own entry among the training and synthetic datasets, rather than added to the model count. | 2026-11-08 | 3 | benchmark | MMLU/GSM8K/IFEval | Tulu-3-405B MMLU 87.0, GSM8K 95.5, IFEval 86.0, HumanEval 95.9, MATH 67.3 | null | null | high | Tulu 3 is a 2025 post-training recipe on a Llama 3.1 base. Its 405B results table gives MMLU 87.0, GSM8K 95.5, IFEval 86.0, HumanEval 95.9 and MATH 67.3, against a GPT-4o (11-24) column at 87.9, 91.7, 84.8, 97.0 and 68.8, and it sits below the current frontier. | 2026-09-29 | 2.4 | adoption_and_capability | null | false | null | [
"tulu-3"
] | null | null | null | null | null | null | null | [
"https://github.com/allenai/open-instruct"
] | null | null | null | null | [
"https://huggingface.co/allenai/Llama-3.1-Tulu-3-8B",
"https://huggingface.co/allenai/Llama-3.1-Tulu-3-70B",
"https://huggingface.co/allenai/Llama-3.1-Tulu-3-405B"
] | [
"https://huggingface.co/datasets/allenai/tulu-3-sft-mixture"
] | null | 10 |
zephyr | Zephyr | model | Hugging Face H4's open chat model line. The flagship zephyr-7b-beta is a DPO fine-tune of Mistral 7B trained with the open alignment-handbook recipe on the public UltraChat and UltraFeedback datasets, making the whole run reproducible. The line also includes a 141B mixture-of-experts ORPO variant. | hugging-face | Hugging Face | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 13 | 5 | open_source | open | weights:open(MIT;base Mistral-7B Apache-2.0);data:open(UltraChat_200k + UltraFeedback_binarized, both MIT, ungated);code:open(alignment-handbook DPO recipe, Apache-2.0);license:MIT(OSI);orpo-checkpoint:open(zephyr-orpo-141b-A35b-v0.1, a separately released ORPO checkpoint, ships under Apache-2.0; this score describes z... | high | Genuinely fully open: MIT weights, both DPO training datasets public and MIT-licensed, and end-to-end training code (the alignment handbook) under Apache-2.0 that reproduces zephyr-7b-beta. One caveat that does not change the class: the upstream UltraChat and UltraFeedback sets were partly generated with OpenAI model o... | 2026-09-29 | 2 | 10K-100K | usage_volume | high | Adoption is measured as Hugging Face downloads of the two declared Zephyr models, almost all of it for the 7B flagship; the 141B ORPO variant sees negligible use. The UltraChat and UltraFeedback alignment datasets the product also declares are not added to the model figure. | 2026-10-25 | 2 | benchmark | MT-Bench/AlpacaEval | zephyr-7b-beta MT-Bench 7.34, AlpacaEval 90.6% win rate, Open LLM Leaderboard avg 52.15 | null | null | high | Zephyr-7B-beta was strong for a 7B model at its late-2023 release, with its card reporting strong results on several MT-Bench categories against larger open models such as Llama2-Chat-70B. It is weak by current standards (GSM8K 12.74, DROP 9.66 on the leaderboard) and sits well below current frontier chat models. | 2026-09-29 | 2 | adoption_and_capability | null | false | null | null | null | null | null | null | null | null | null | [
"https://github.com/huggingface/alignment-handbook"
] | null | null | null | null | [
"https://huggingface.co/HuggingFaceH4/zephyr-7b-beta",
"https://huggingface.co/HuggingFaceH4/zephyr-orpo-141b-A35b-v0.1"
] | [
"https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k",
"https://huggingface.co/datasets/HuggingFaceH4/ultrafeedback_binarized"
] | null | 12 |
k2 | K2 | model | LLM360 and MBZUAI's K2 model line. K2-V2, a 70B dense model released in December 2025 as a reasoning-enhanced LLM the project calls 360-open, ships its weights alongside the complete TxT360 pretraining, mid-training and SFT data, the training code and intermediate checkpoints, so the run is reproducible; it superseded ... | llm360 | LLM360 | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 14 | 5 | open_source | open | weights:open(Apache-2.0);data:open(TxT360 pretraining + midtraining + SFT);code:open(k2v2_train);checkpoints:open;license:Apache-2.0(OSI) | high | K2-V2 ships Apache-2.0 weights, the complete TxT360 pretraining, midtraining and SFT data, training code and checkpoints, a fully reproducible open release at 70B. | 2026-09-29 | 1 | <10K | usage_volume | high | Adoption is measured as Hugging Face downloads across the two declared checkpoints, the base and instruct models. Both see light use. | 2026-11-08 | 4 | benchmark | Artificial Analysis Intelligence Index | K2 Horizon 375B-A23B scores 31 on the Artificial Analysis Intelligence Index, ranked #14 of 117 open-weights entries (median 18); its model card reports GPQA Diamond 87.3, Terminal-Bench 2.1 70.2 and SWE Bench Pro 42.6. | gemini-pro | at | medium | This score reads K2 Horizon, the line's current flagship, which ships under Apache-2.0 with its pretraining and mid-training data. It is among the stronger open-weights models on the Artificial Analysis index and level there with Gemini 3.1 Pro, but well behind the closed leading group, so it scores level with Gemini P... | 2026-10-08 | 3.1 | adoption_and_capability | null | false | null | [
"k2-v2"
] | null | null | null | null | null | null | null | [
"https://github.com/LLM360/k2v2_train"
] | null | null | null | null | [
"https://huggingface.co/IFM/K2-V2",
"https://huggingface.co/IFM/K2-V2-Instruct"
] | null | null | 7 |
phi-instruct | Phi-4-mini-instruct | model | Phi-4-mini-instruct is a 3.8B-parameter dense decoder-only language model with a 128K-token context window, tuned for chat and instruction following through supervised fine-tuning and direct preference optimization. Microsoft positions it for memory-constrained and latency-bound settings. It belongs to the Phi-4 family... | microsoft | Microsoft | null | finetuned_chat | Fine-tuned / chat models | model_components | Models | 15 | 3 | open_weights | open-ish | weights:open(MIT, on HF);data:closed(synthetic-heavy training data undisclosed);code:closed(no training pipeline);license:MIT(OSI permissive) | high | MIT-licensed open weights, with the training and post-training data and the training code unreleased, so this is open weights rather than open source. | 2026-09-29 | 3 | 100K-1M | usage_volume | high | Adoption is read from Hugging Face downloads on the one declared artifact, microsoft/Phi-4-mini-instruct. | 2026-11-08 | 2 | benchmark | MMLU-Pro/GPQA | MMLU 67.3, MMLU-Pro 52.8, GPQA 25.2, GSM8K 88.6, MATH 64.0 (overall 63.5) | null | null | high | Phi-4-mini-instruct is strong on math for a 3.8B model (GSM8K 88.6) but weak on graduate-level science questions (GPQA 25.2), and it trails GPT-4o-mini overall on the model card's quality table. It sits well below current frontier chat models. | 2026-09-29 | 2.3 | adoption_and_capability | null | false | null | [
"phi-4-mini-instruct"
] | null | null | null | null | null | null | null | null | null | null | null | null | [
"https://huggingface.co/microsoft/Phi-4-mini-instruct"
] | null | null | 5 |
End of preview. Expand in Data Studio
Open Source AI Gap Map
The data behind the Open Source AI Gap Map: 1,269 products from 701 organizations across 35 categories of the AI stack, each graded on openness, adoption and capability, with every score backed by cited evidence.
- Version: 0.6.0, released 2026-11-12
- Source: currentai-org/os-ai-map at commit
938e8e51230c0bffb2ad4c9ecc3def80235d3a44 - Method: methodology (source)
from datasets import load_dataset
products = load_dataset("current-ai-official/os-ai-map", "products", split="train")
citations = load_dataset("current-ai-official/os-ai-map", "citations", split="train")
# A specific release:
products_v = load_dataset("current-ai-official/os-ai-map", "products", revision="v0.6.0", split="train")
Tables
| Config | Grain | Rows |
|---|---|---|
products |
one published product | 1,269 |
organizations |
one organization that owns or stewards a published product | 701 |
categories |
one published category | 35 |
citations |
one cited source behind one product's axis score | 9,501 |
product_history |
one product in one release | 4,776 |
category_history |
one category in one release | 142 |
Join keys: products.slug = citations.product_slug = product_history.current_slug;
products.organization_slug = organizations.slug; products.category_slug = categories.slug.
In the history tables product_slug is the slug as that release published it; join on
current_slug to follow a renamed product to its row in products.
products
Each product sits in exactly one category and one organization. Columns are grouped by axis:
- Openness (
openness_*): a 0–5 grade against openness frameworks (the Model Openness Framework for models, OSI license classes for software).openness_classis the named class (open_source,open_weights,restricted, ...) andopenness_bucketcollapses it toopen,open-ishorclosed.openness_componentslists the license, access and documentation findings behind the grade. - Adoption (
adoption_*): a 0–5 level from observed usage where it can be measured (downloads, active users, deployments). Where no usage instrument exists the level rests on a proxy, andadoption_signal_typesays which:stars_fallback(GitHub stars, capped at level 3 because stars measure attention rather than use),reported_traction(figures the vendor reports), orunknown.adoption_reachis the band the level rests on, such as1M-10M. Raw measurements are not published here. - Capability (
capability_*): a 0–5 grade from community benchmarks where they exist and feature coverage where they don't.capability_relative_toandcapability_relationname the peer product the grade was set against. - Overall (
overall_score,overall_basis,tier,mature): adoption and capability blended with the category's weights.tierisleading(≥ 4.5),strong(4.0–4.5) or null.overall_basissays whether a score blends both axes or rests on adoption alone; don't compare the two kinds directly.
Every axis carries a *_confidence, a *_note explaining the grade in prose, and a
*_last_verified date: the last day the whole axis was confirmed against its sources. It is not
derived from the citations' accessed dates. A recently fetched page does not mean the score was
reassessed that day, and a score can be reconfirmed without a page being refetched. Null means
the axis has not yet been confirmed under that rule. Artifact columns (github, pypi, huggingface_model, ...) list the
URLs the product is identified and measured by. lineage_* columns name the products (by slug)
or external works a product derives from, trains, or belongs to.
categories
stage_num / stage_name is the category's Maturity Stage, from 0 (Void) to 5 (Mature), and
gaps names what the open ecosystem still lacks there (adoption, capability, openness,
...). Both are derived from the category's products; the rules are in the methodology.
citations
The audit trail. Each row is one source cited for one axis of one product: the url, what the
page shows (a short quotation or paraphrase), when it was accessed, the http_status seen,
and a content_sha256 of the fetched page so a reader can tell whether it has changed since.
establishes names the facts the source was cited for, where recorded.
Citations are the evidence a score was checked against. They include primary sources (license
files, package registries, model cards, papers) and supporting third-party evidence, so a
citation is not a guarantee that the source is primary. accessed dates the fetch, not the
score; see *_last_verified above.
product_history and category_history
The map as built at each release, from v0.1.0 onward (6 snapshots in this export). Each snapshot is the build output committed at the release tag, and the export checks it against a fresh rebuild of that tag's own sources with that tag's own code. A snapshot that differs from its rebuild fails the export.
Rows carry the release version and released date, the source_commit the release tag points
at, and the payload_commit that wrote the build output, which can be earlier than the tag. The
history block of manifest.json records how each snapshot was verified. Some releases' build
output still carries the previous version number, because the release commit itself does not
trigger a rebuild. The manifest records that, and the content is verified the same way.
These are snapshots of what each release contained, not of when the live site showed it. The site is published in a separate step that leaves no record in the repository.
Read changes between releases with care. A value can move because the product changed, because new evidence arrived, or because the method changed: categories were added and split, adoption bands were re-drawn, and the taxonomy was revised. The repository's CHANGELOG records each method change against the release that made it. Coverage also grows release over release, so a product missing from an early release was usually not yet scored rather than absent.
What is not included
- The long tail. The map counts tens of thousands of discovered repositories, models and packages it has not scored. That inventory draws on third-party catalogs and is left out.
- Preliminary categories. Categories still being populated are not on the published map and are not here.
- Raw adoption measurements. Download counts and other observations live in the warehouse; this dataset carries the curated level and band.
Limitations
Scores are curated judgments against a published rubric, made by people and reviewed agents, and checked against cited sources. They are not measurements. Closed products are included deliberately, so open options can be compared against them. Adoption is hard to observe for some kinds of product (hosted services, hardware, apps), and the note on each score says what its instrument can and cannot show.
Corrections are welcome as
issues or
pull requests against sources/ in the source repository.
License
MIT, the same as the source repository. Cited pages remain the property of their owners; the
shows field quotes them briefly as evidence.
Citation
@misc{osaimap,
title = {Open Source AI Gap Map},
author = {{Current AI} and {os-ai-map contributors}},
year = {2026},
note = {Version 0.6.0},
url = {https://github.com/currentai-org/os-ai-map}
}
- Downloads last month
- 43