🚨⚠️ I HAVE REACHED HUGGING FACE'S FREE STORAGE LIMIT ⚠️🚨
I can no longer upload new models unless I can cover the cost of additional storage. I host 70+ free models as an independent contributor and this work is unpaid. Without your support, no more new models can be uploaded.
Your help will motivate me and would go into further improving my workflow and coverings fees for storage, compute and may even help uncensoring bigger model with rental Cloud GPUs.
This is an Experts-Only NVFP4 quantization of Qwen3.5-35B-A3B. MoE expert tensors are quantized with NVFP4; non-expert/sensitive tensors are preserved at higher precision. The displayed Hub metadata may be missing because the MoE export contains a very large number of physical tensors, this NVFP4 safetensors contain BF16, F32, U8, and F8_E4M3 tensors.
Lower refusals indicate fewer content restrictions, while lower KL divergence indicates better preservation of the original model's capabilities. Higher refusals cause more rejections, objections, pushbacks, lecturing, censorship, softening and deflections, while higher KL divergence degrades coherence, reasoning ability, and overall quality.
[!Note]
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.
[!Tip]
For users seeking managed, scalable inference without infrastructure maintenance, the official Qwen API service is provided by Alibaba Cloud Model Studio.
In particular, Qwen3.5-Flash is the hosted version corresponding to Qwen3.5-35B-A3B with more production features, e.g., 1M context length by default and official built-in tools.
For more information, please refer to the User Guide.
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.
Benchmark Results
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: 35B in total and 3B activated
Number of Linear Attention Heads: 32 for V and 16 for QK
Head Dimension: 128
Gated Attention:
Number of Attention Heads: 16 for Q and 2 for KV
Head Dimension: 256
Rotary Position Embedding Dimension: 64
Mixture Of Experts
Number of Experts: 256
Number of Activated Experts: 8 Routed + 1 Shared
Expert Intermediate Dimension: 512
LM Output: 248320 (Padded)
MTP: trained with multi-steps
Context Length: 262,144 natively and extensible up to 1,010,000 tokens.
Benchmark Results
Language
GPT-5-mini 2025-08-07
GPT-OSS-120B
Qwen3-235B-A22B
Qwen3.5-122B-A10B
Qwen3.5-27B
Qwen3.5-35B-A3B
Knowledge
MMLU-Pro
83.7
80.8
84.4
86.7
86.1
85.3
MMLU-Redux
93.7
91.0
93.8
94.0
93.2
93.3
C-Eval
82.2
76.2
92.1
91.9
90.5
90.2
SuperGPQA
58.6
54.6
64.9
67.1
65.6
63.4
Instruction Following
IFEval
93.9
88.9
87.8
93.4
95.0
91.9
IFBench
75.4
69.0
51.7
76.1
76.5
70.2
MultiChallenge
59.0
45.3
50.2
61.5
60.8
60.0
Long Context
AA-LCR
68.0
50.7
60.0
66.9
66.1
58.5
LongBench v2
56.8
48.2
54.8
60.2
60.6
59.0
STEM & Reasoning
HLE w/ CoT
19.4
14.9
18.2
25.3
24.3
22.4
GPQA Diamond
82.8
80.1
81.1
86.6
85.5
84.2
HMMT Feb 25
89.2
90.0
85.1
91.4
92.0
89.0
HMMT Nov 25
84.2
90.0
89.5
90.3
89.8
89.2
Coding
SWE-bench Verified
72.0
62.0
--
72.0
72.4
69.2
Terminal Bench 2
31.9
18.7
--
49.4
41.6
40.5
LiveCodeBench v6
80.5
82.7
75.1
78.9
80.7
74.6
CodeForces
2160
2157
2146
2100
1899
2028
OJBench
40.4
41.5
32.7
39.5
40.1
36.0
FullStackBench en
30.6
58.9
61.1
62.6
60.1
58.1
FullStackBench zh
35.2
60.4
63.1
58.7
57.4
55.0
General Agent
BFCL-V4
55.5
--
54.8
72.2
68.5
67.3
TAU2-Bench
69.8
--
58.5
79.5
79.0
81.2
VITA-Bench
13.9
--
31.6
33.6
41.9
31.9
DeepPlanning
17.9
--
17.1
24.1
22.6
22.8
Search Agent
HLE w/ tool
35.8
19.0
--
47.5
48.5
47.4
Browsecomp
48.1
41.1
--
63.8
61.0
61.0
Browsecomp-zh
49.5
42.9
--
69.9
62.1
69.5
WideSearch
47.2
40.4
--
60.5
61.1
57.1
Seal-0
34.2
45.1
--
44.1
47.2
41.4
Multilingualism
MMMLU
86.2
78.2
83.4
86.7
85.9
85.2
MMLU-ProX
78.5
74.5
77.9
82.2
82.2
81.0
NOVA-63
51.9
51.1
55.4
58.6
58.1
57.1
INCLUDE
81.8
74.0
81.0
82.8
81.6
79.7
Global PIQA
88.5
84.1
85.7
88.4
87.5
86.6
PolyMATH
67.3
54.0
60.1
68.9
71.2
64.4
WMT24++
80.7
74.4
75.8
78.3
77.6
76.3
MAXIFE
85.3
83.7
83.2
87.9
88.0
86.6
* CodeForces: evaluated on our own query set.
* 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.
* Search Agent: most search agents built on our model adopt a simple context-folding strategy(256k): once the cumulative Tool Response length reaches a preset threshold, earlier Tool Responses are pruned from the history to keep the context within limits.
* WideSearch: we use a 256k context window without any context management.
* 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).
* Empty cells (--) indicate scores not yet available or not applicable.
Vision Language
GPT-5-mini 2025-08-07
Claude-Sonnet-4.5
Qwen3-VL-235B-A22B
Qwen3.5-122B-A10B
Qwen3.5-27B
Qwen3.5-35B-A3B
STEM and Puzzle
MMMU
79.0
79.6
80.6
83.9
82.3
81.4
MMMU-Pro
67.3
68.4
69.3
76.9
75.0
75.1
MathVision
71.9
71.1
74.6
86.2
86.0
83.9
Mathvista(mini)
79.1
79.8
85.8
87.4
87.8
86.2
DynaMath
81.4
78.8
82.8
85.9
87.7
85.0
ZEROBench
3
4
4
9
10
8
ZEROBench_sub
27.3
26.3
28.4
36.2
36.2
34.1
VlmsAreBlind
75.8
85.5
79.5
96.7
96.9
97.0
BabyVision
20.9
18.6
22.2
40.2 / 34.5
44.6 / 34.8
38.4 / 29.6
General VQA
RealWorldQA
79.0
70.3
81.3
85.1
83.7
84.1
MMStar
74.1
73.8
78.7
82.9
81.0
81.9
MMBenchEN-DEV-v1.1
86.8
88.3
89.7
92.8
92.6
91.5
SimpleVQA
56.8
57.6
61.3
61.7
56.0
58.3
HallusionBench
63.2
59.9
66.7
67.6
70.0
67.9
Text Recognition and Document Understanding
OmniDocBench1.5
77.0
85.8
84.5
89.8
88.9
89.3
CharXiv(RQ)
68.6
67.2
66.1
77.2
79.5
77.5
MMLongBench-Doc
50.3
--
56.2
59.0
60.2
59.5
CC-OCR
70.8
68.1
81.5
81.8
81.0
80.7
AI2D_TEST
88.2
87.0
89.2
93.3
92.9
92.6
OCRBench
82.1
76.6
87.5
92.1
89.4
91.0
Spatial Intelligence
ERQA
54.0
45.0
52.5
62.0
60.5
64.8
CountBench
91.0
90.0
93.7
97.0
97.8
97.8
RefCOCO(avg)
--
--
91.1
91.3
90.9
89.2
ODInW13
--
--
43.2
44.5
41.1
42.6
EmbSpatialBench
80.7
71.8
84.3
83.9
84.5
83.1
RefSpatialBench
9.0
2.2
69.9
69.3
67.7
63.5
LingoQA
62.4
12.8
66.8
80.8
82.0
79.2
Hypersim
--
--
11.0
12.7
13.0
13.1
SUNRGBD
--
--
34.9
36.2
35.4
33.4
Nuscene
--
--
13.9
15.4
15.2
14.6
Video Understanding
VideoMME(w sub.)
83.5
81.1
83.8
87.3
87.0
86.6
VideoMME(w/o sub.)
78.9
75.3
79.0
83.9
82.8
82.5
VideoMMMU
82.5
77.6
80.0
82.0
82.3
80.4
MLVU
83.3
72.8
83.8
87.3
85.9
85.6
MVBench
--
--
75.2
76.6
74.6
74.8
LVBench
--
--
63.6
74.4
73.6
71.4
MMVU
69.8
70.6
71.1
74.7
73.3
72.3
Visual Agent
ScreenSpot Pro
--
36.2
62.0
70.4
70.3
68.6
OSWorld-Verified
--
61.4
38.1
58.0
56.2
54.5
AndroidWorld
--
--
63.7
66.4
64.2
71.1
Tool Calling
TIR-Bench
24.6
27.6
29.8
53.2 / 42.5
59.8 / 42.3
55.5 / 38.0
V*
71.7
58.6
85.9
93.2 / 90.1
93.7 / 89.0
92.7 / 89.5
Medical VQA
SLAKE
70.5
73.6
54.7
81.6
80.0
78.7
PMC-VQA
36.3
55.9
41.2
63.3
62.4
62.0
MedXpertQA-MM
34.4
54.0
47.6
67.3
62.4
61.4
* MathVision: our model’s score is evaluated using a fixed prompt, e.g., “Please reason step by step, and put your final answer within \boxed{}.” For other models, we report the higher score between runs with and without the \boxed{} formatting.
* BabyVision: scores reported as "with CI / without CI".
* TIR-Bench and V*: scores reported as "with CI / without CI".
* Empty cells (--) indicate scores not yet available or not applicable.
Quickstart
[!Important]
Qwen3.5 models operate in thinking mode by default, generating thinking content signified by <think>\n...</think>\n\n before producing the final responses.
To disable thinking content and obtain direct response, 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.
[!Important]
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.
[!Important]
The model has a default context length of 262,144 tokens.
If you encounter out-of-memory (OOM) errors, consider reducing the context window.
However, because Qwen3.5 leverages extended context for complex tasks, we advise maintaining a context length of at least 128K tokens to preserve thinking capabilities.
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:
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 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:
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.
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:
1from openai import OpenAI
2# Configured by environment variables3client = OpenAI()45messages =[6{7"role":"user",8"content":[9{10"type":"image_url",11"image_url":{12"url":"https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/CI_Demo/mathv-1327.jpg"13}14},15{16"type":"text",17"text":"The centres of the four illustrated circles are in the corners of the square. The two big circles touch each other and also the two little circles. With which factor do you have to multiply the radii of the little circles to obtain the radius of the big circles?\nChoices:\n(A) $\\frac{2}{9}$\n(B) $\\sqrt{5}$\n(C) $0.8 \\cdot \\pi$\n(D) 2.5\n(E) $1+\\sqrt{2}$"18}19]20}21]2223response = client.chat.completions.create(24 model="Qwen/Qwen3.5-35B-A3B",25 messages=messages,26 max_tokens=81920,27 temperature=1.0,28 top_p=0.95,29 presence_penalty=1.5,30 extra_body={31"top_k":20,32},33)34print("Chat response:", chat_response)
Video Input
python
1from openai import OpenAI
2# Configured by environment variables3client = OpenAI()45messages =[6{7"role":"user",8"content":[9{10"type":"video_url",11"video_url":{12"url":"https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/video/N1cdUjctpG8.mp4"13}14},15{16"type":"text",17"text":"How many porcelain jars were discovered in the niches located in the primary chamber of the tomb?"18}19]20}21]2223# When vLLM is launched with `--media-io-kwargs '{"video": {"num_frames": -1}}'`,24# video frame sampling can be configured via `extra_body` (e.g., by setting `fps`).25# This feature is currently supported only in vLLM.26#27# By default, `fps=2` and `do_sample_frames=True`.28# With `do_sample_frames=True`, you can customize the `fps` value to set your desired video sampling rate.29response = client.chat.completions.create(30 model="Qwen/Qwen3.5-35B-A3B",31 messages=messages,32 max_tokens=81920,33 temperature=1.0,34 top_p=0.95,35 presence_penalty=1.5,36 extra_body={37"top_k":20,38"mm_processor_kwargs":{"fps":2,"do_sample_frames":True},39},40)4142print("Chat response:", chat_response)
Instruct (or Non-Thinking) Mode
[!Important]
Qwen3.5 does not officially support the soft switch of Qwen3, i.e., /think and /nothink.
Qwen3.5 will think by default before response.
You can obtain direct response from the model without thinking by configuring the API parameters.
For example,
[!Note]
If you are using APIs from Alibaba Cloud Model Studio, in addition to changing model, please use "enable_thinking": False instead of "chat_template_kwargs": {"enable_thinking": False}.
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.
python
1import os
2from qwen_agent.agents import Assistant
34# Define LLM5# Using Alibaba Cloud Model Studio6llm_cfg ={7# Use the OpenAI-compatible model service provided by DashScope:8'model':'Qwen3.5-35B-A3B',9'model_type':'qwenvl_oai',10'model_server':'https://dashscope.aliyuncs.com/compatible-mode/v1',11'api_key': os.getenv('DASHSCOPE_API_KEY'),1213'generate_cfg':{14'use_raw_api':True,15# When using Dash Scope OAI API, pass the parameter of whether to enable thinking mode in this way16'extra_body':{17'enable_thinking':True18},19},20}2122# Using OpenAI-compatible API endpoint.23# functionality of the deployment frameworks and let Qwen-Agent automate the related operations.24#25# llm_cfg = {26# # Use your own model service compatible with OpenAI API by vLLM/SGLang:27# 'model': 'Qwen/Qwen3.5-35B-A3B',28# 'model_type': 'qwenvl_oai',29# 'model_server': 'http://localhost:8000/v1', # api_base30# 'api_key': 'EMPTY',31#32# 'generate_cfg': {33# 'use_raw_api': True,34# # When using vLLM/SGLang OAI API, pass the parameter of whether to enable thinking mode in this way35# 'extra_body': {36# 'chat_template_kwargs': {'enable_thinking': True}37# },38# },39# }4041# Define Tools42tools =[43{'mcpServers':{# You can specify the MCP configuration file44"filesystem":{45"command":"npx",46"args":["-y","@modelcontextprotocol/server-filesystem","/Users/xxxx/Desktop"]47}48}49}50]5152# Define Agent53bot = Assistant(llm=llm_cfg, function_list=tools)5455# Streaming generation56messages =[{'role':'user','content':'Help me organize my desktop.'}]57for responses in bot.run(messages=messages):58pass59print(responses)6061# Streaming generation62messages =[{'role':'user','content':'Develop a dog website and save it on the desktop'}]63for responses in bot.run(messages=messages):64pass65print(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.
Qwen3.5 natively supports context lengths of up to 262,144 tokens.
For long-horizon tasks where the total length (including both input and output) exceeds this limit, we recommend using RoPE scaling techniques to handle long texts effectively., e.g., YaRN.
YaRN is currently supported by several inference frameworks, e.g., transformers, vllm, ktransformers and sglang.
In general, there are two approaches to enabling YaRN for supported frameworks:
Modifying the model configuration file:
In the config.json file, change the rope_parameters fields in text_config to:
[!NOTE]
All the notable open-source frameworks implement static YaRN, which means the scaling factor remains constant regardless of input length, potentially impacting performance on shorter texts.
We advise modifying the rope_parameters configuration only when processing long contexts is required.
It is also recommended to modify the factor as needed. For example, if the typical context length for your application is 524,288 tokens, it would be better to set factor as 2.0.
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:
Thinking mode for general tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
For supported frameworks, you can adjust the presence_penalty parameter 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 answer field 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 size parameter in the released video_preprocessor_config.json is conservatively configured. It is recommended to set the longest_edge parameter 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,
