Instructions to use LTM-Lab/Qwen3.5-0.8B-v14 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use LTM-Lab/Qwen3.5-0.8B-v14 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-text-to-text", model="LTM-Lab/Qwen3.5-0.8B-v14", trust_remote_code=True) messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] pipe(text=messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModelForImageTextToText processor = AutoProcessor.from_pretrained("LTM-Lab/Qwen3.5-0.8B-v14", trust_remote_code=True) model = AutoModelForImageTextToText.from_pretrained("LTM-Lab/Qwen3.5-0.8B-v14", trust_remote_code=True, device_map="auto") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=256) print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use LTM-Lab/Qwen3.5-0.8B-v14 with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "LTM-Lab/Qwen3.5-0.8B-v14" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "LTM-Lab/Qwen3.5-0.8B-v14", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/LTM-Lab/Qwen3.5-0.8B-v14
- SGLang
How to use LTM-Lab/Qwen3.5-0.8B-v14 with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "LTM-Lab/Qwen3.5-0.8B-v14" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "LTM-Lab/Qwen3.5-0.8B-v14", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "LTM-Lab/Qwen3.5-0.8B-v14" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "LTM-Lab/Qwen3.5-0.8B-v14", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }' - Docker Model Runner
How to use LTM-Lab/Qwen3.5-0.8B-v14 with Docker Model Runner:
docker model run hf.co/LTM-Lab/Qwen3.5-0.8B-v14
Qwen3.5-0.8B-v14
็น้ซไธญๆๆ่ฆ๏ผ้ๆฏ Qwen/Qwen3.5-0.8B ็ V14 ่ฝๆ็ใๆฌ้่ๅๆจกๅ้ไฝๅ
็ธๅ๏ผๅค็ๆฏ่จๆถ็่จญๅฎ่ๅ
ฉๅ็จๅผๆชใ็จ Docker ้กๅ lovewin10/sglang-v14:nightly-dev-cu13-20260924-ffac53d7-v14๏ผไฟฎๆน้็ SGLang๏ผๆๅๆ๏ผๆจกๅๅธถๆ่ทจ่ซๆฑไฟๅญ็ๆ
็ฏ่จๆถ๏ผๅจไธๆฎตๅฐ่ฉฑ่ฃก่ชช้็ไบ๏ผไนๅพๅจๅฆไธๆฎตๅ
จๆฐ็ๅฐ่ฉฑ่ฃกๅๅพๅฐ๏ผCUDA graphใๅ็ถดๅฟซๅ่ๆจๆธฌ่งฃ็ขผ็
งๅธธ้่ใๅฏฆๆธฌ็ตๆ่ๅทฒ็ฅ็้ๅถๅฏซๅจไธ้ขใ
Qwen/Qwen3.5-0.8B converted to the V14 episodic memory architecture (Qwen3_5V14ForConditionalGeneration).
The weights are the original's, byte for byte. Nothing was trained, merged or quantized here.
Added: the architecture name, an
auto_mapand the memory options (qwen_dense_v14_memory_config) inconfig.json, and two code files (configuration_qwen_dense_v14_reinstatement.py,modeling_qwen_dense_v14_reinstatement.py) that hold the memory.Served by the Docker image
lovewin10/sglang-v14:nightly-dev-cu13-20260924-ffac53d7-v14, a modified SGLang, the model has a memory that persists across requests: what was said in one chat is answered in a later, separate chat. The store is on disk and survives restarts.The small model of the family: the one the memory was developed and benchmarked on. Use it to try the mechanism; what it recalls it uses less reliably than the larger models do.
How the memory works
While a prompt is read, the keys and values its attention layers compute are cut into events of up to 128 tokens and written to a store on disk, each with a low-rank engram of the state it leaves in the model's recurrent layers. A later prompt is read once to see what it asks for; the stored events whose keys would take its attention are recalled by one vote for all layers and heads; their engrams are composed into the request's recurrent state and the prompt is read again with the recalled keys and values merged into the attention. Nothing is retrieved as text and nothing is added to the prompt. A question that asks for nothing in the store recalls little or nothing, but its answer is not always the plain model's word for word: a stored event that resembles it can pass the vote, and while generating the request also attends the event the memory's world model expects next (foresight).
Serve with SGLang (Docker)
Needs any GPU; the Transformers class also runs on CPU, Docker with the NVIDIA container runtime, and a writable directory for the memory store.
hf download LTM-Lab/Qwen3.5-0.8B-v14 --local-dir <models>/Qwen3.5-0.8B-v14
docker run --rm -it --gpus all --ipc=host -p 8000:8000 `
-v "<models>:/model:ro" `
-v "<memory store>:/memory" `
lovewin10/sglang-v14:nightly-dev-cu13-20260924-ffac53d7-v14 `
sglang serve --model-path /model/Qwen3.5-0.8B-v14 --trust-remote-code `
--host 0.0.0.0 --port 8000 --mem-fraction-static 0.78 --served-model-name qwen-0.8b-v14 `
--chat-template /model/Qwen3.5-0.8B-v14/chat_template.jinja `
--reasoning-parser qwen3-thinking --tool-call-parser qwen3_coder `
--attention-backend flashinfer --linear-attn-backend triton --mamba-ssm-dtype float32 `
--speculative-algorithm EAGLE --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 `
--enable-linear-replayssm-spec --mamba-radix-cache-strategy extra_buffer_lazy `
--context-length 262144 --max-prefill-tokens 32768 --max-running-requests 8 `
--kv-cache-dtype fp8_e4m3 --max-mamba-cache-size 16
(PowerShell line continuations; in a POSIX shell end the lines with \.) The server speaks the OpenAI API on port 8000. The image is lmsysorg/sglang:nightly-dev-cu13-20260924-ffac53d7 with its SGLang replaced; a launch command written for that image runs unchanged apart from the /memory mount.
/memoryholds what the prompts contained: the token ids and the attention keys and values of every prompt the server has read. Treat it like chat logs; delete the directory to forget everything.- Every request reads and writes the same store. There is no separation between users: run one server with its own
/memoryfor each user or tenant whose memories must stay apart. - Generated text is not written. A reply becomes memory when a later prompt carries it, as chat front-ends do when they resend the conversation.
-e SGLANG_V14_MEMORY=0serves the directory without memory, as the original model.-e SGLANG_V14_TRACE=1logs every recall decision (the votes, the events recalled).-e SGLANG_V14_INDEX_GB(default 2) is the GPU memory for the index's rows. With more stored payloads than fit, a vote reads the rest back from the store (seconds a vote from a hard disk) and the log says so once: raise it, or keep the store on an SSD.-e SGLANG_V14_WRITE_QUEUE_GB(default 8) bounds the RAM of events waiting to be written; a request beyond it stores nothing more, and the log says so.- The index keeps a little per event in memory and the rest on disk, so the server's memory stays level as the store grows; every 10 minutes the log says what it holds, the process's own memory apart from shared memory. Images pushed before the evening of 2026-11-12 kept about 0.2 MB of host memory for every event they stored and, after a world model fit, the learner's copy of the index until a later garbage collection: pull the image again. A copy of this directory downloaded before 2026-11-12 has an older
modeling_qwen_dense_v14_reinstatement.pywhose index keeps every event in memory (the server warns once): download that file again. - The memory needs
--attention-backend flashinfer(orfa3), and TP, PP and DP attention of 1.
Use with Transformers
The two code files are remote code for Transformers; the SGLang image runs the same code for every memory decision.
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor
path = 'LTM-Lab/Qwen3.5-0.8B-v14'
model = AutoModelForImageTextToText.from_pretrained(path, trust_remote_code=True, dtype=torch.bfloat16).to('cuda').eval()
processor = AutoProcessor.from_pretrained(path)
def ask(text):
messages = [{'role': 'user', 'content': [{'type': 'text', 'text': text}]}]
prompt = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False, enable_thinking=False)
ids = processor.tokenizer(prompt, return_tensors='pt', add_special_tokens=False).input_ids.to(model.device)
out = model.generate(input_ids=ids, max_new_tokens=40, do_sample=False)
return processor.tokenizer.decode(out[0, ids.shape[1]:], skip_special_tokens=True)
ask('Please remember: my locker code is 4817-2953.')
ask('What is my locker code?') # a separate call: the first one is recalled from the store
Every forward reads and writes the store on disk (under the user's state directory, named after the weights; the memory option storage_path sets another place). model(...).memory_trace lists what each row wrote and recalled, model.open_memory_bank() returns the store (memories(), forget(name)), and model.memory_enabled = False gives the original model. The benchmark below was run this way.
Measured
The memory benchmark of the project (examples/memory_benchmark.py) on these weights with the Transformers class: bf16, greedy decoding, 40 new tokens per answer. Each section writes earlier sessions into an empty store and then asks in a new session; a hit is a reply that contains the reference answer. Three to ten questions per section, so a difference of one or two means nothing; a rerun of the same code moves a section by one or two.
| Section | With the memory | Everything in one context, no memory (upper bound) | Question only |
|---|---|---|---|
| single: ten facts, a store each | 9 / 10 | 8 / 10 | 0 |
| noise: ten facts among 48 other sessions | 7 / 10 (the event with the answer recalled for 10) | 8 / 10 | 0 |
| crowd: ten facts among 248 other sessions | 6 / 10 (recalled for 8) | ||
| update: a fact corrected in a later session, asked for its present value | 1 / 6 (1 more answered the old value; 1 to 4 over the runs) | 6 / 6 | |
| hop: two facts from two sessions needed for one answer | 6 / 6 | 6 / 6 | |
| long: a fact in the middle of a long document | 3 / 3 | 3 / 3 | |
| unrelated questions (the noise store open) | 7 / 8, no event recalled | 7 / 8 without memory |
Served by the SGLang image the directory behaves as the larger models' directories do; the image's own checks were run on the 27B and the Nemotron conversions (their cards).
Limits
- It recalls what a question asks for in similar words. The vote compares the question's attention queries with stored keys. A question that shares little with the note that answers it can miss it; the larger the model, the less this happens.
- Two-step questions are the weak case. "In which rack is the backup server?" needs the note that names the server and then the note about that server. The second is recalled only if its vote reaches half of the first one's (
recall_ratio), and often it falls just short. - A corrected fact is not retired. The old note and the new one are both in the store; when both are recalled the model usually, not always, answers the newer one. Recalled events carry no time stamp in this image.
- The first token waits for the recall. A prompt that recalls something is read two or three times. Decoding is somewhat slower while recalled events are attended.
- A request waits until what the server read before it is stored, so that it can recall it: unnoticeable in chat, up to some ten seconds right after a document of several thousand tokens.
- No separation between users, and the store keeps what prompts contained (see above).
- Disk. About 6 KB per prompt token, plus a state engram per event. Events unused for seven days are rewritten in a 4-bit format.
- Recalling is not answering: among 48 other sessions the note with the answer was recalled for ten of ten questions and seven were answered; among 248 it was recalled for eight and six were answered.
- A corrected fact is the unsteady case on this model: one to four of six over five runs of the benchmark.
- The recall thresholds (
recall_ratio,recall_floorin the memory options) were set on Qwen3.5-0.8B and not re-tuned for this model. - The trained weights are untouched, so everything the original card says about the model itself still holds, including its own limitations.
Credits and license
- The model is Qwen/Qwen3.5-0.8B; its card follows below, unchanged, and its license applies to this repository (see the metadata above and the original card).
- The serving engine is SGLang (Apache-2.0); the image replaces the SGLang of
lmsysorg/sglang:nightly-dev-cu13-20260924-ffac53d7and is not an official SGLang image. - Conversion, memory code and image: LTM-Lab / lovewin10.
Original model card (Qwen/Qwen3.5-0.8B)
Qwen3.5-0.8B
This repository contains model weights and configuration files for the post-trained model in the Hugging Face Transformers format.
These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, KTransformers, etc.
In light of its parameter scale, the intended use cases are prototyping, task-specific fine-tuning, and other research or development purposes.
Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance. Qwen3.5 represents a significant leap forward, integrating breakthroughs in multimodal learning, architectural efficiency, reinforcement learning scale, and global accessibility to empower developers and enterprises with unprecedented capability and efficiency.
Qwen3.5 Highlights
Qwen3.5 features the following enhancement:
Unified Vision-Language Foundation: Early fusion training on multimodal tokens achieves cross-generational parity with Qwen3 and outperforms Qwen3-VL models across reasoning, coding, agents, and visual understanding benchmarks.
Efficient Hybrid Architecture: Gated Delta Networks combined with sparse Mixture-of-Experts deliver high-throughput inference with minimal latency and cost overhead.
Scalable RL Generalization: Reinforcement learning scaled across million-agent environments with progressively complex task distributions for robust real-world adaptability.
Global Linguistic Coverage: Expanded support to 201 languages and dialects, enabling inclusive, worldwide deployment with nuanced cultural and regional understanding.
Next-Generation Training Infrastructure: Near-100% multimodal training efficiency compared to text-only training and asynchronous RL frameworks supporting massive-scale agent scaffolds and environment orchestration.
For more details, please refer to our blog post Qwen3.5.
Model Overview
- Type: Causal Language Model with Vision Encoder
- Training Stage: Pre-training & Post-training
- Language Model
- Number of Parameters: 0.8B
- Hidden Dimension: 1024
- Token Embedding: 248320 (Padded)
- Number of Layers: 24
- Hidden Layout: 6 ร (3 ร (Gated DeltaNet โ FFN) โ 1 ร (Gated Attention โ FFN))
- Gated DeltaNet:
- Number of Linear Attention Heads: 16 for V and 16 for QK
- Head Dimension: 128
- Gated Attention:
- Number of Attention Heads: 8 for Q and 2 for KV
- Head Dimension: 256
- Rotary Position Embedding Dimension: 64
- Feed Forward Network:
- Intermediate Dimension: 3584
- LM Output: 248320 (Tied to token embedding)
- MTP: trained with multi-steps
- Context Length: 262,144 natively
Benchmark Results
Language
| Qwen3-4B-2507 | Qwen3-1.7B | Qwen3.5-2B | Qwen3.5-0.8B | |
|---|---|---|---|---|
| Non-Thinking Mode | ||||
| MMLU-Pro | 69.6 | 40.2 | 55.3 | 29.7 |
| MMLU-Redux | 84.2 | 64.4 | 69.2 | 48.5 |
| C-Eval | 80.2 | 61.0 | 65.2 | 46.4 |
| SuperGPQA | 42.8 | 21.0 | 30.4 | 16.9 |
| IFEval | 83.4 | 68.2 | 61.2 | 52.1 |
| MMMLU | 64.9 | 46.7 | 56.9 | 34.1 |
| Knowledge & STEM (Thinking) | ||||
| MMLU-Pro | 74.0 | 56.5 | 66.5 | 42.3 |
| MMLU-Redux | 86.1 | 73.9 | 79.6 | 59.5 |
| C-Eval | 82.2 | 68.1 | 73.2 | 50.5 |
| SuperGPQA | 47.8 | 31.2 | 37.5 | 21.3 |
| GPQA | 65.8 | 40.1 | 51.6 | 11.9 |
| Instruction Following (Thinking๏ผ | ||||
| IFEval | 87.4 | 72.5 | 78.6 | 44.0 |
| IFBench | 50.4 | 26.7 | 41.3 | 21.0 |
| MultiChallenge | 41.7 | 27.2 | 33.7 | 18.9 |
| Long Context (Thinking๏ผ | ||||
| AA-LCR | 32.0 | 6.7 | 25.6 | 4.7 |
| LongBench v2 | 42.8 | 26.5 | 38.7 | 26.1 |
| Reasoning (Thinking๏ผ | ||||
| HMMT Feb 25 | 57.5 | 10.2 | 22.9 | -- |
| HMMT Nov 25 | 69.6 | 8.9 | 19.6 | -- |
| General Agent (Thinking๏ผ | ||||
| BFCL-V4 | 39.9 | -- | 43.6 | 25.3 |
| TAU2-Bench | 43.2 | -- | 48.8 | 11.6 |
| Multilingualism (Thinking๏ผ | ||||
| MMMLU | 70.8 | 57.0 | 63.1 | 44.3 |
| MMLU-ProX | 62.4 | 49.4 | 52.3 | 34.6 |
| NOVA-63 | 47.1 | 40.3 | 46.4 | 42.4 |
| INCLUDE | 64.4 | 51.8 | 55.4 | 40.6 |
| Global PIQA | 73.5 | 63.1 | 69.3 | 59.4 |
| PolyMATH | 46.2 | 25.2 | 26.1 | 8.2 |
| WMT24++ | 58.9 | 39.3 | 45.8 | 27.2 |
| MAXIFE | 72.1 | 50.7 | 60.6 | 39.2 |
* TAU2-Bench: we follow the official setup except for the airline domain, where all models are evaluated by applying the fixes proposed in the Claude Opus 4.5 system card.
* MMLU-ProX: we report the averaged accuracy on 29 languages.
* WMT24++: a harder subset of WMT24 after difficulty labeling and rebalancing; we report the averaged scores on 55 languages using XCOMET-XXL.
* MAXIFE: we report the accuracy on English + multilingual original prompts (totally 23 settings).
* Experimental settings: top_p=0.95, top_k=20, presence_penalty=1.5, and temperature=1.0 were used.
* Empty cells (--) indicate scores not yet available or not applicable.
Vision Language
| Qwen3-VL-4B | Qwen3-VL-2B | Qwen3.5-2B | Qwen3.5-0.8B | |
|---|---|---|---|---|
| STEM and Puzzle | ||||
| MMMU | 70.8 | 61.4 | 64.2/64.2 | 49/47.4 |
| MMMU-Pro | 57.0 | 42.5 | 50.3/47.7 | 31.2/31.4 |
| Mathvista(mini) | 79.5 | 73.6 | 76.7/73.9 | 62.2/58.6 |
| DynaMath | 74.4 | 66.7 | 73.6/69.6 | 49.9/46.5 |
| ZEROBench | 0.0 | 0.0 | 1.0/0.0 | 0.0/0.0 |
| ZEROBench_sub | 18.9 | 13.2 | 17.1/18.6 | 12.9/11.4 |
| VlmsAreBlind | 68.6 | 50.0 | 75.8/74.3 | 59.4/57.3 |
| General VQA | ||||
| RealWorldQA | 73.2 | 69.5 | 74.5/71.2 | 63.4/61.6 |
| MMStar | 73.2 | 68.1 | 71.7/68.0 | 58.3/55.9 |
| MMBenchEN-DEV-v1.1 | 86.7 | 81.9 | 83.3/81.3 | 69.9/68.0 |
| SimpleVQA | 48.8 | 43.6 | 38.5/39.5 | 31.3/30.4 |
| HallusionBench | 64.1 | 54.9 | 58.0/51.3 | 53.1/46.7 |
| Text Recognition and Document Understanding | ||||
| MMLongBench-Doc | 44.4 | 33.8 | 45.4/38.8 | 33.6/28.1 |
| AI2D_TEST | 84.9 | 80.4 | 83.3/81.5 | 69.9/68.7 |
| CC-OCR | 73.8 | 68.3 | 72.9/75.8 | 63.2/66.7 |
| OmniDocBench1.5 | 80.0 | 65.9 | 79.8/80.9 | 61.0/70.6 |
| CharXiv(RQ) | 50.3 | 37.1 | 58.8/52.6 | 41.3/38.2 |
| OCRBench | 80.8 | 79.2 | 84.5/85.4 | 74.5/79.1 |
| Spatial Intelligence | ||||
| RefCOCO(avg) | 88.2 | 84.8 | 84.8/84.3 | 79.3/77.8 |
| CountBench | 89.4 | 84.1 | 91.4/86.8 | 77.0/68.6 |
| ODInW13 | 39.4 | 36.0 | 35.9/40.5 | 31.6/33.2 |
| ERQA | 47.3 | 41.8 | 43.8/33.0 | 34.5/23.8 |
| EmbSpatialBench | 80.7 | 75.9 | 77.9/66.4 | 68.6/54.6 |
| RefSpatialBench | 45.3 | 28.9 | 32.9/30.0 | 23.5/21.7 |
| Hypersim | 11.9 | 11.2 | 12.4/12.4 | 11.9/11.0 |
| SUNRGBD | 28.0 | 28.6 | 28.7/25.6 | 26.1/23.3 |
| Nuscene | 4.9 | 4.0 | 6.9/8.5 | 5.7/7.0 |
| Video Understanding | ||||
| VideoMME(w sub.) | 76.0 | 67.9 | 75.6/-- | 63.8/-- |
| VideoMME(w/o sub.) | 68.9 | 62.1 | 69.0/-- | 57.7/-- |
| VideoMMMU | 69.4 | 54.1 | 62.1/-- | 44.3/-- |
| MLVU | 75.7 | 69.2 | 76.2/-- | 65.6/-- |
| MVBench | 69.3 | 64.5 | 64.9/-- | 55.8/-- |
| LVBench | 53.5 | 47.6 | 57.1/-- | 45.1/-- |
| MMVU | 58.6 | 48.9 | 48.6/-- | 34.3/-- |
| Visual Agent | ||||
| ScreenSpot Pro | 59.5 | 48.5 | --/54.5 | --/46.5 |
| Medical VQA | ||||
| SLAKE | 65.9 | 61.1 | 74.4/67.5 | 62.6/59.5 |
| PMC-VQA | 48.4 | 42.4 | 48.8/54.0 | 40.4/45.5 |
| MedXpertQA-MM | 26.3 | 13.0 | 26.9/19.1 | 17.1/25.3 |
* Scores of Qwen3.5 models are reported as Thinking / Non-thinking.
* MathVision: our modelโs score is evaluated using a fixed prompt, e.g., โPlease reason step by step, and put your final answer within \boxed{}.โ For other models, we report the higher score between runs with and without the \boxed{} formatting.
* Experimental settings: For the Video benchmarks, we used top_p=0.95, top_k=20, presence_penalty=1.5, and temperature=1.0. All other benchmarks adopted the same sampling configuration but with temperature=0.6 under the thinking mode. Under the non-thinking mode, the sampling parameters were set to top_p=0.8, top_k=20, presence_penalty=1.5, and temperature=0.7.
* Empty cells (--) indicate scores not yet available or not applicable.
Quickstart
Qwen3.5 models support both non-thinking and thinking mode. Qwen3.5-0.8B operates in non-thinking mode by default. To enable thinking, refer to the examples here.
For streamlined integration, we recommend using Qwen3.5 via APIs. Below is a guide to use Qwen3.5 via OpenAI-compatible API.
Serving Qwen3.5
Qwen3.5 can be served via APIs with popular inference frameworks. In the following, we show example commands to launch OpenAI-Compatible API servers for Qwen3.5 models.
Inference efficiency and throughput vary significantly across frameworks. We recommend using the latest framework versions to ensure optimal performance and compatibility. For production workloads or high-throughput scenarios, dedicated serving engines such as SGLang, KTransformers or vLLM are strongly recommended.
The model has a default context length of 262,144 tokens. If you encounter out-of-memory (OOM) errors, consider reducing the context window.
SGLang
SGLang is a fast serving framework for large language models and vision language models. SGLang from the main branch of the open-source repository is required for Qwen3.5, which can be installed using the following command in a fresh environment:
uv pip install 'git+https://github.com/sgl-project/sglang.git#subdirectory=python&egg=sglang[all]'
See its documentation for more details.
The following will create API endpoints at http://localhost:8000/v1:
Standard Version: The following command can be used to create an API endpoint with maximum context length 262,144 tokens using tensor parallel on 8 GPUs.
python -m sglang.launch_server --model-path Qwen/Qwen3.5-0.8B --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144Tool Use: To support tool use, you can use the following command.
python -m sglang.launch_server --model-path Qwen/Qwen3.5-0.8B --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144 --tool-call-parser qwen3_coderMulti-Token Prediction (MTP): The following command is recommended for MTP:
python -m sglang.launch_server --model-path Qwen/Qwen3.5-0.8B --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144 --speculative-algo NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4
vLLM
vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. vLLM from the main branch of the open-source repository is required for Qwen3.5, which can be installed using the following command in a fresh environment:
uv pip install vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly
See its documentation for more details.
For detailed Qwen3.5 usage guide, see the vLLM Qwen3.5 recipe.
The following will create API endpoints at http://localhost:8000/v1:
Standard Version: The following command can be used to create an API endpoint with maximum context length 262,144 tokens using tensor parallel on 8 GPUs.
vllm serve Qwen/Qwen3.5-0.8B --port 8000 --tensor-parallel-size 1 --max-model-len 262144Tool Call: To support tool use, you can use the following command.
vllm serve Qwen/Qwen3.5-0.8B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser qwen3_coderMulti-Token Prediction (MTP): The following command is recommended for MTP:
vllm serve Qwen/Qwen3.5-0.8B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'Text-Only: The following command skips the vision encoder and multimodal profiling to free up memory for additional KV cache:
vllm serve Qwen/Qwen3.5-0.8B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --language-model-only
KTransformers
KTransformers is a flexible framework for experiencing cutting-edge LLM inference optimizations with CPU-GPU heterogeneous computing. For running Qwen3.5 with KTransformers, see the KTransformers Deployment Guide.
Hugging Face Transformers
Hugging Face Transformers contains a lightweight server which can be used for quick testing and moderate load deployment.
The latest transformers is required for Qwen3.5:
pip install "transformers[serving] @ git+https://github.com/huggingface/transformers.git@main"
See its documentation for more details. Please also make sure torchvision and pillow are installed.
Then, run transformers serve to launch a server with API endpoints at http://localhost:8000/v1; it will place the model on accelerators if available:
transformers serve --force-model Qwen/Qwen3.5-0.8B --port 8000 --continuous-batching
Using Qwen3.5 via the Chat Completions API
The chat completions API is accessible via standard HTTP requests or OpenAI SDKs. Here, we show examples using the OpenAI Python SDK.
Before starting, make sure it is installed and the API key and the API base URL is configured, e.g.:
pip install -U openai
# Set the following accordingly
export OPENAI_BASE_URL="http://localhost:8000/v1"
export OPENAI_API_KEY="EMPTY"
We recommend using the following set of sampling parameters for generation
- Non-thinking mode for text tasks:
temperature=1.0, top_p=1.00, top_k=20, min_p=0.0, presence_penalty=2.0, repetition_penalty=1.0- Non-thinking mode for VL tasks:
temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0- Thinking mode for text tasks:
temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0- Thinking mode for VL or precise coding (e.g. WebDev) tasks :
temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0Please note that the support for sampling parameters varies according to inference frameworks.
Text-Only Input
from openai import OpenAI
# Configured by environment variables
client = OpenAI()
messages = [
{"role": "user", "content": "Give me a short introduction to large language models."},
]
chat_response = client.chat.completions.create(
model="Qwen/Qwen3.5-0.8B",
messages=messages,
max_tokens=32768,
temperature=1.0,
top_p=1.0,
presence_penalty=2.0,
extra_body={
"top_k": 20,
},
)
print("Chat response:", chat_response)
Image Input
from openai import OpenAI
# Configured by environment variables
client = OpenAI()
messages = [
{
"role": "user",
"content": [
{
"type": "image_url",
"image_url": {
"url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/RealWorld/RealWorld-04.png"
}
},
{
"type": "text",
"text": "Where is this?"
}
]
}
]
chat_response = client.chat.completions.create(
model="Qwen/Qwen3.5-0.8B",
messages=messages,
max_tokens=32768,
temperature=0.7,
top_p=0.8,
presence_penalty=1.5,
extra_body={
"top_k": 20,
},
)
print("Chat response:", chat_response)
Video Input
from openai import OpenAI
# Configured by environment variables
client = OpenAI()
messages = [
{
"role": "user",
"content": [
{
"type": "video_url",
"video_url": {
"url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/video/N1cdUjctpG8.mp4"
}
},
{
"type": "text",
"text": "Summarize the video content."
}
]
}
]
# When vLLM is launched with `--media-io-kwargs '{"video": {"num_frames": -1}}'`,
# video frame sampling can be configured via `extra_body` (e.g., by setting `fps`).
# This feature is currently supported only in vLLM.
#
# By default, `fps=2` and `do_sample_frames=True`.
# With `do_sample_frames=True`, you can customize the `fps` value to set your desired video sampling rate.
chat_response = client.chat.completions.create(
model="Qwen/Qwen3.5-0.8B",
messages=messages,
max_tokens=32768,
temperature=0.7,
top_p=0.8,
presence_penalty=1.5,
extra_body={
"top_k": 20,
"mm_processor_kwargs": {"fps": 2, "do_sample_frames": True},
},
)
print("Chat response:", chat_response)
Thinking Mode
Qwen3.5 does not officially support the soft switch of Qwen3, i.e.,
/thinkand/nothink.
You can make the model think before response by configuring the API parameters. For example,
from openai import OpenAI
# Configured by environment variables
client = OpenAI()
messages = [
{"role": "user", "content": "Type \"I love Qwen3.5\" backwards"},
]
chat_response = client.chat.completions.create(
model="Qwen/Qwen3.5-0.8B",
messages=messages,
max_tokens=81920,
temperature=1.0,
top_p=0.95,
presence_penalty=1.5,
extra_body={
"top_k": 20,
"enable_thinking": True,
},
)
print("Chat response:", chat_response)
In thinking mode, we have observed that when using the recommended sampling parameters, Qwen3.5-0.8B is more prone to entering thinking loops compared to other Qwen3.5 models, which may prevent it from terminating generation properly. We recommend further tuning the sampling parameters specific to your use case and utilizing the API's streaming generation mode (if supported) to enable timely detection and interruption of such anomalous generation behaviors.
Agentic Usage
Qwen3.5 excels in tool calling capabilities.
Qwen-Agent
We recommend using Qwen-Agent to quickly build Agent applications with Qwen3.5.
To define the available tools, you can use the MCP configuration file, use the integrated tool of Qwen-Agent, or integrate other tools by yourself.
import os
from qwen_agent.agents import Assistant
# Define LLM
# Using OpenAI-compatible API endpoint. The API backend should disable response parsers.
llm_cfg = {
# Use your own model service compatible with OpenAI API by vLLM/SGLang:
'model': 'Qwen/Qwen3.5-0.8B',
'model_type': 'qwenvl_oai',
'model_server': 'http://localhost:8000/v1', # api_base
'api_key': 'EMPTY',
'generate_cfg': {
'use_raw_api': True,
# Pass the parameter of whether to enable thinking mode in this way
# 'extra_body': {
# 'chat_template_kwargs': {'enable_thinking': True}
# },
},
}
# Define Tools
tools = [
{'mcpServers': { # You can specify the MCP configuration file
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/xxxx/Desktop"]
}
}
}
]
# Define Agent
bot = Assistant(llm=llm_cfg, function_list=tools)
# Streaming generation
messages = [{'role': 'user', 'content': 'Help me organize my desktop.'}]
for responses in bot.run(messages=messages):
pass
print(responses)
# Streaming generation
messages = [{'role': 'user', 'content': 'Develop a dog website and save it on the desktop'}]
for responses in bot.run(messages=messages):
pass
print(responses)
Qwen Code
Qwen Code is an open-source AI agent for the terminal, optimized for Qwen models. It helps you understand large codebases, automate tedious work, and ship faster.
For more information, please refer to Qwen Code.
Best Practices
To achieve optimal performance, we recommend the following settings:
Sampling Parameters:
We suggest using the following sets of sampling parameters depending on the mode and task type:
- Non-thinking mode for text tasks:
temperature=1.0,top_p=1.00,top_k=20,min_p=0.0,presence_penalty=2.0,repetition_penalty=1.0 - Non-thinking mode for VL tasks:
temperature=0.7,top_p=0.80,top_k=20,min_p=0.0,presence_penalty=1.5,repetition_penalty=1.0 - Thinking mode for text tasks:
temperature=1.0,top_p=0.95,top_k=20,min_p=0.0,presence_penalty=1.5,repetition_penalty=1.0 - Thinking mode for VL or precise coding (e.g., WebDev) tasks:
temperature=0.6,top_p=0.95,top_k=20,min_p=0.0,presence_penalty=0.0,repetition_penalty=1.0
- Non-thinking mode for text tasks:
For supported frameworks, you can adjust the
presence_penaltyparameter between 0 and 2 to reduce endless repetitions. However, using a higher value may occasionally result in language mixing and a slight decrease in model performance.
Adequate Output Length: We recommend using an output length of 32,768 tokens for most queries. For benchmarking on highly complex problems, such as those found in math and programming competitions, we suggest setting the max output length to 81,920 tokens. This provides the model with sufficient space to generate detailed and comprehensive responses, thereby enhancing its overall performance.
Standardize Output Format: We recommend using prompts to standardize model outputs when benchmarking.
- Math Problems: Include "Please reason step by step, and put your final answer within \boxed{}." in the prompt.
- Multiple-Choice Questions: Add the following JSON structure to the prompt to standardize responses: "Please show your choice in the
answerfield with only the choice letter, e.g.,"answer": "C"."
No Thinking Content in History: In multi-turn conversations, the historical model output should only include the final output part and does not need to include the thinking content. It is implemented in the provided chat template in Jinja2. However, for frameworks that do not directly use the Jinja2 chat template, it is up to the developers to ensure that the best practice is followed.
Long Video Understanding: To optimize inference efficiency for plain text and images, the
sizeparameter in the releasedvideo_preprocessor_config.jsonis conservatively configured. It is recommended to set thelongest_edgeparameter in the video_preprocessor_config file to 469,762,048 (corresponding to 224k video tokens) to enable higher frame-rate sampling for hour-scale videos and thereby achieve superior performance. For example,{"longest_edge": 469762048, "shortest_edge": 4096}Alternatively, override the default values via engine startup parameters. For implementation details, refer to: vLLM / SGLang.
Citation
If you find our work helpful, feel free to give us a cite.
@misc{qwen3.5,
title = {{Qwen3.5}: Towards Native Multimodal Agents},
author = {{Qwen Team}},
month = {February},
year = {2026},
url = {https://qwen.ai/blog?id=qwen3.5}
}
- Downloads last month
- 79