World models & exotic LLMs
Frontier and off-beat releases I track — language world models, persona merges and experimental abliterated fine-tunes.
Text Generation • 35B • Updated • 9.13k • 748Note Native language WORLD MODEL — simulates agent environments. A simulator, not an agent.
Qwen/Qwen3.8-Flash-Next
Image-Text-to-Text • 180B • Updated • 2.98M • 7.26kNote Frontier exhibit : archi declaree Qwen4ExpForConditionalGeneration / model_type qwen4_exp — une archi Qwen4 publiee sous un nom Qwen3.8. 180B, hors 3090 (ollama ne publie que 125b-a6b nvfp4/mlx) : expose, pas installe.
-
zai-org/GLM-5.3-Flash
Image-Text-to-Text • 321B • Updated • 6.39M • • 3.25k
zai-org/GLM-5.2
Text Generation • 753B • Updated • 304k • • 5.21kNote Z.ai flagship MoE — frontier open-weights release
-
deepseek-ai/DeepSeek-V4.1-Flash
Image-Text-to-Text • 763B • Updated • 1.61M • • 4.99k
thinkingmachines/Inkling
Image-Text-to-Text • 952B • Updated • 1.45M • • 1.86kNote Thinking Machines' first open release — ~1T-param multimodal MoE (image+audio→text). Far beyond local hardware; here as a frontier exhibit.
thinkingmachines/Inkling-2-Small
Image-Text-to-Text • 594B • Updated • 18.4k • • 438Note 383B MoE follow-up to Inkling, Apache-2.0, served by deepinfra and together. Exhibit only, far beyond a 24 GB card.
ricursiveAI/Ripple-65B-Preview
65B • Updated • 13.1k • 612Note 65B dense (not MoE) transformer, 12T tokens, 1M context, Apache-2.0, Olmo 3 lineage. Vendor-reported agentic-coding scores, not independently checked.
openpangu/openPangu-2.0-Flash
100B • Updated • 288 • 53Note Huawei openPangu 100B — custom architecture, frontier release outside the usual labs
NaiveAI/Naive-N0.5-Flash
Text Generation • Updated • 15.8k • 412Note 309B MoE (15.5B active), custom architecture, MIT: a lab pitching 'frontier AI built with AI'. Exhibit only, far beyond one 3090.
danielrmay/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-W4A16
Text Generation • 81B • Updated • 93 • 8Note Nemotron 3 Puzzle 75B-A9B — latent-MoE + Mamba hybrid, NAS-assembled architecture
-
prism-ml/Ternary-Bonsai-2-27B-gguf
Text Generation • 27B • Updated • 1.41M • 3.04k
mateogrgic/GLM-5.2-colibri-int4-with-int8-mtp
Updated • 419 • 144Note GLM-5.2 'colibri' int4 + int8 MTP — expert-streaming to run a giant MoE on CPU
MaralGPT/MaralGPT-Mythos-9B-2606-GGUF
Text Generation • 9B • Updated • 2.84k • 67Note Persona merge — voice/character experiment.
DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-THINKING-HERETIC-UNCENSORED
Image-Text-to-Text • 9B • Updated • 872 • 247Note Exotic high-IQ thinking merge, uncensored.
AliesTaha/fable-traces
Text Generation • 4B • Updated • 103 • 210Note Qwen3-4B fine-tuned on Claude Fable 5 traces — frontier behavior distilled into 4B
ConceptCountry/qwen-concept-27b
27B • Updated • 65.1k • 367Note Qwen3.8-27B with nine down-projections rewritten by weight-space programming, no training. Reports fewer thinking tokens and higher GPQA, measured at Q8_0 only. Experimental.
-
unsloth/Ornith-1.5-122B-GGUF
Text Generation • 125B • Updated • 831k • 107 -
bartowski/DeepSeek-V4-Flash-GGUF
Text Generation • 284B • Updated • 356 • 36 -
ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF
Text Generation • 197B • Updated • 4.32M • 676 -
ConwayResearch/Underdog-Saluki-27B-1.0
Text Generation • 27B • Updated • 482k • 719 -
orcarouter/OrcaSAQ-2-27B
Text Generation • 7B • Updated • 1.3k • 288 -
sudoingx/Ternary-Bonsai-2-27B-PTQ1_0-MTP-GGUF
Text Generation • 27B • Updated • 22.3k • 75 -
turboderp/Qwen3.8-27B-exl3
Updated • 5.05k • 209 -
darkmaniac7/Qwen3.5-9B-uncensored-MNN
Text Generation • Updated • 794 • 12 -
Jiunsong/SuperHY3-abliterated-gguf
Text Generation • 299B • Updated • 61 • 6 -
AnkitAI/Parable-Qwen3-4B-Claude-Fable-5
Text Generation • 4B • Updated • 90 • 6