This model's benchmarks are in the OpenAI, Claude and Gemini "zone" - it is that strong.
All Optimized Quants BENCHMARKED (5 metrics): You know exactly how strong each NEO-CODE quant is relative to full precision model.
IQ4XS stands out at 94% of full precision at 1/4 the size of the full model. Then there are the Q6 and Q8, with the latter hitting 98.38% (of full precision). Even
the lowest (IQ2_M) at 1/5 (10.5GB) the size of the full model scores 82.66%! All quants and metrics listed below.
Uncensored quants are here (exceeds all metrics of the quants at this repo too): Click here
Qwen3.6-27B-NEO-CODE-Di-IMatrix-MAX-GGUF
Team Qwen exceeded all expectations with this 27B model [even exceeding their own 398B model] AND GEMMA 4s too, so here are the quants to match.
All GGUFs benchmarked against full precision model (also primer below) and there is an Unsloth VS DavidAU GGUF showdown too. (You guys know I love the Unsloth team.)
Quant "engineering" focused on balance and precision, vs raw power (which seemed in some cases to destabilize the model/quant).
In other words benchmarks / stats determined the best quants, not guesswork or one size fits all approach.
This was done to ensure long context, long/multi-convos, coding and math etc etc performed as close as possible to full precision model
as well as one-shot, and standard prompting / problem solving.
TWO Imatrix datasets were used to do this by first getting "raw stats" on both, then merging them to get the best of each imatrix in one dataset
then this was used to make the "NEO-CODE-Di-IMatrix-MAX" quants.
Additional tensor adjustments were also made, which were also measured (benched) and adjusted too.
Benchmarks from Team Qwen below in the Qwen Model section.
HINT: They are jaw dropping.
These quants are 256K max context.
Enjoy!
Uncensored version?
Fully uncensored, and the model was fine tuned and NEO-Code DiMatrix quants too:
GGUF BENCHMARKS relative to Full Precision: (primer below to explain "numbers")
All quants listed below are Imatrix'ed using NEO-CODE di-matrix.
Metric
IQ2_M
IQ3_M
IQ4_XS
IQ4_NL
Q4_K_S
Q4_K_M
Q5_K_S
Q5_K_M
Q6_K
Q8_0
Same Top P (%)
82.66%
89.63%
93.98%
94.04%
93.90%
94.33%
95.84 %
96.09%
97.34%
98.38%
Mean KLD
0.1840
0.0749
0.0261
0.0260
0.0267
0.0242
0.0142
0.0132
0.0056
0.0034
99.9% KLD
7.22
4.88
2.18
2.36
2.34
2.62
1.69
1.58
0.50
0.20
RMS Δp (%)
12.52%
7.69%
4.34%
4.36%
4.41%
4.13%
3.22%
3.019%
1.988%
1.538%
Mean PPL (Q)
7.746
7.222
6.977
6.971
6.948
6.946
(!) 6.894
(!) 6.885
6.924
6.914
NOTES:
With the exception of "Same Top P (%)" (how close it matchs full precision), lower is better for all other metrics.
"(!)" for "Mean PPL (Q)" are LOWER than the BF16/full (6.900) precision model.
Q8_0 contains BF16 components, and not affected by IMATRIX. This is an ULTIMATE PERFORMANCE quant. I have also put the full breakdown of metrics in detail for this quant below.
Q2s/Q3s are not here because IQ2/3s are faster, and smaller with same or slightly better quality.
Summary of Results: (primer below to explain)
Pure Power: Q6_K leads the pack; and comes within 2.66% of BF16 precision; but Q8_0 beats every quant (but with a much larger footprint memory wise).
Peak Intelligence: Q5_K_S owns the high ground in almost every category in terms of leap from Q4s to Q5s ; but there is not a lot of difference between Q5s and Q4s/IQ4s too.
Maximum Reliability: IQ4_XS remains the winner [in 4 bit quants] for the 99.9% KLD (2.18), meaning it is the least likely to have extreme mathematical glitches in the "long tail" of tokens.
The IQ Winner: IQ4_NL is a fantastic middle ground if you want high accuracy (94.04%) with a slightly smaller file size than the K-Quants.
UNSLOTH VS DAVIDAU Showdown: imatrix'ed Q4_K_S (primer below to explain the numbers)
Notes:
"dataset #4 (Quant #5)" refers to Unsloth's Q4_K_S (of Qwen3.6-27B) quant at their repo (imatrix).
"hybrid dataset" / "Final Q4_K_S" refers to my NEO-CODE Di-Matrix (imatrix) Quant Q4_K_S.
All NEO-CODE Di-Matrix quants follow the same formula to maintain quality as shown in NEO-CODE Di-Matrix (imatrix) Quant Q4_K_S below.
Metric
Dataset #4 (Quant #5)
Final Q4_K_S (Hybrid #1)
Winner
Same Top P (%)
93.92%
93.90%
Dataset #4 (by 0.02%)
Mean KLD
0.0263
0.0267
Dataset #4
99.9% KLD (Stability)
2.75
2.34
Final Q4_K_S
Maximum KLD (Outliers)
23.90
21.97
Final Q4_K_S
RMS Δp (%)
4.38%
4.41%
Dataset #4
Mean PPL (Q)
6.966
6.948
Final Q4_K_S
Key Takeaways:
The Intelligence Tie: Dataset #4 has a tiny lead in Same Top P (93.92% vs 93.90%), but it's effectively a tie in real-world usage.
The Reliability Win: Your Final Q4_K_S is much safer for long-context or creative work. The 99.9% KLD (2.34) and Max KLD (21.97) prove that your hybrid dataset + final tensor adjustments successfully "tamed" the instability seen in the raw Dataset #4 run.
Superior Fluency: The lower Perplexity (6.948) of the Final Q4_K_S means the model's overall language distribution is more natural and less "noisy."
A Beginner's Primer to Quantization Metrics
Quantization compresses Large Language Models to make them run faster and on cheaper hardware. To know if a model is still "smart" after compression, we use these five key metrics:
1. Same Top P (%)
What it is: How often the compressed model picks the exact same word as its first choice compared to the original, uncompressed model.
In plain English: This is "Word-for-Word Accuracy." If this is 94%, it means in 94 out of 100 cases, the model’s top choice remains identical to the original.
The Goal: Higher is better (93% or above is near-perfect).
2. Mean KLD (KL Divergence)
What it is: A measure of how much the "logic" or "thought process" of the model has drifted. It looks at the probabilities of all possible next words, not just the top one.
In plain English: This is the "Reasoning Loss." It measures how much the model's internal "brain" changed during compression.
The Goal: Lower is better (Below 0.03 is excellent).
3. 99.9% KLD (Stability)
What it is: This focuses on the "worst" 0.1% of tokens—the most difficult edge cases the model encountered during testing.
In plain English: This is the "Reliability Score." It tells you if the model is prone to "glitching" or producing gibberish when the conversation gets complicated.
The Goal: Lower is better (Lower numbers mean a more stable model).
4. RMS Δp (%)
What it is: The average change in the model’s confidence levels.
In plain English: This is "Confidence Alignment." Even if the model picks the right word, does it feel as sure as the original? High numbers mean the model feels "jittery" or hesitant.
The Goal: Lower is better (Ideally near 4%).
5. Mean PPL (Perplexity)
What it is: A measure of how "surprised" the model is by the text it is reading.
In plain English: This is "Fluency." If perplexity goes up significantly, the model’s writing will feel less natural, more robotic, or repetitive.
The Goal: Lower is better (Should be as close to the Base model as possible).
[!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.
Following the February release of the Qwen3.5 series, we're pleased to share the first open-weight variant of Qwen3.6. Built on direct feedback from the community, Qwen3.6 prioritizes stability and real-world utility, offering developers a more intuitive, responsive, and genuinely productive coding experience.
Qwen3.6 Highlights
This release delivers substantial upgrades, particularly in
Agentic Coding: the model now handles frontend workflows and repository-level reasoning with greater fluency and precision.
Thinking Preservation: we've introduced a new option to retain reasoning context from historical messages, streamlining iterative development and reducing overhead.
Benchmark Results
For more details, please refer to our blog post Qwen3.6-27B.
Number of Linear Attention Heads: 48 for V and 16 for QK
Head Dimension: 128
Gated Attention:
Number of Attention Heads: 24 for Q and 4 for KV
Head Dimension: 256
Rotary Position Embedding Dimension: 64
Feed Forward Network:
Intermediate Dimension: 17408
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
Qwen3.5-27B
Qwen3.5-397B-A17B
Gemma4-31B
Claude 4.5 Opus
Qwen3.6-35B-A3B
Qwen3.6-27B
Coding Agent
SWE-bench Verified
75.0
76.2
52.0
80.9
73.4
77.2
SWE-bench Pro
51.2
50.9
35.7
57.1
49.5
53.5
SWE-bench Multilingual
69.3
69.3
51.7
77.5
67.2
71.3
Terminal-Bench 2.0
41.6
52.5
42.9
59.3
51.5
59.3
SkillsBench Avg5
27.2
30.0
23.6
45.3
28.7
48.2
QwenWebBench
1068
1186
1197
1536
1397
1487
NL2Repo
27.3
32.2
15.5
43.2
29.4
36.2
Claw-Eval Avg
64.3
70.7
48.5
76.6
68.7
72.4
Claw-Eval Pass^3
46.2
48.1
25.0
59.6
50.0
60.6
QwenClawBench
52.2
51.8
41.7
52.3
52.6
53.4
Knowledge
MMLU-Pro
86.1
87.8
85.2
89.5
85.2
86.2
MMLU-Redux
93.2
94.9
93.7
95.6
93.3
93.5
SuperGPQA
65.6
70.4
65.7
70.6
64.7
66.0
C-Eval
90.5
93.0
82.6
92.2
90.0
91.4
STEM & Reasoning
GPQA Diamond
85.5
88.4
84.3
87.0
86.0
87.8
HLE
24.3
28.7
19.5
30.8
21.4
24.0
LiveCodeBench v6
80.7
83.6
80.0
84.8
80.4
83.9
HMMT Feb 25
92.0
94.8
88.7
92.9
90.7
93.8
HMMT Nov 25
89.8
92.7
87.5
93.3
89.1
90.7
HMMT Feb 26
84.3
87.9
77.2
85.3
83.6
84.3
IMOAnswerBench
79.9
80.9
74.5
84.0
78.9
80.8
AIME26
92.6
93.3
89.2
95.1
92.7
94.1
* SWE-Bench Series: Internal agent scaffold (bash + file-edit tools); temp=1.0, top_p=0.95, 200K context window. We correct some problematic tasks in the public set of SWE-bench Pro and evaluate all baselines on the refined benchmark.
* Terminal-Bench 2.0: Harbor/Terminus-2 harness; 3h timeout, 32 CPU/48 GB RAM; temp=1.0, top_p=0.95, top_k=20, max_tokens=80K, 256K ctx; avg of 5 runs.
* SkillsBench: Evaluated via OpenCode on 78 tasks (self-contained subset, excluding API-dependent tasks); avg of 5 runs.
* NL2Repo: Others are evaluated via Claude Code (temp=1.0, top_p=0.95, max_turns=900).
* QwenClawBench: A real-user-distribution Claw agent benchmark; temp=0.6, 256K ctx.
* QwenWebBench: An internal front-end code generation benchmark; bilingual (EN/CN), 7 categories (Web Design, Web Apps, Games, SVG, Data Visualization, Animation, and 3D); auto-render + multimodal judge (code/visual correctness); BT/Elo rating system.
* AIME 26: We use the full AIME 2026 (I & II), where the scores may differ from Qwen 3.5 notes.
Vision Language
Qwen3.5-27B
Qwen3.5-397B-A17B
Gemma4-31B
Claude 4.5 Opus
Qwen3.6-35B-A3B
Qwen3.6-27B
STEM & Puzzle
MMMU
82.3
85.0
80.4
80.7
81.7
82.9
MMMU-Pro
75.0
79.0
76.9
70.6
75.3
75.8
MathVista mini
87.8
--
79.3
--
86.4
87.4
DynaMath
87.7
86.3
79.5
79.7
82.8
85.6
VlmsAreBlind
96.9
--
87.2
--
96.6
97.0
General VQA
RealWorldQA
83.7
83.9
72.3
77.0
85.3
84.1
MMStar
81.0
83.8
77.3
73.2
80.7
81.4
MMBenchEN-DEV-v1.1
92.6
--
90.9
--
92.8
92.3
SimpleVQA
56.0
67.1
52.9
65.7
58.9
56.1
Document Understanding
CharXiv RQ
79.5
80.8
67.9
68.5
78.0
78.4
CC-OCR
81.0
82.0
75.7
76.9
81.9
81.2
OCRBench
89.4
--
86.1
--
90.0
89.4
Spatial Intelligence
ERQA
60.5
67.5
57.5
46.8
61.8
62.5
CountBench
97.8
97.2
96.1
90.6
96.1
97.8
RefCOCO avg
90.9
92.3
--
--
92.0
92.5
EmbSpatialBench
84.5
--
--
--
84.3
84.6
RefSpatialBench
67.7
--
4.7
--
64.3
70.0
Video Understanding
VideoMME(w sub.)
87.0
87.5
--
77.7
86.6
87.7
VideoMMMU
82.3
84.7
81.6
84.4
83.7
84.4
MLVU
85.9
86.7
--
81.7
86.2
86.6
MVBench
74.6
77.6
--
67.2
74.6
75.5
Visual Agent
V*
93.7
95.8
--
67.0
90.1
94.7
AndroidWorld
64.2
--
--
--
--
70.3
* Empty cells (--) indicate scores not yet available or not applicable.
Quickstart
For streamlined integration, we recommend using Qwen3.6 via APIs. Below is a guide to use Qwen3.6 via OpenAI-compatible API.
Serving Qwen3.6
Qwen3.6 can be served via APIs with popular inference frameworks.
In the following, we show example commands to launch OpenAI-Compatible API servers for Qwen3.6 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 Qwen3.6 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>=0.5.10 is recommended for Qwen3.6, 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>=0.19.0 is recommended for Qwen3.6, 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 Qwen3.6 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 Qwen3.6:
pip install "transformers[serving]"
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:
Please note that the support for sampling parameters varies according to inference frameworks.
[!Important]
Qwen3.6 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.
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/Qwen3.6-27B",25 messages=messages,26 max_tokens=81920,27 temperature=1.0,28 top_p=0.95,29 presence_penalty=0.0,30 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/Qwen3.6-27B",31 messages=messages,32 max_tokens=81920,33 temperature=1.0,34 top_p=0.95,35 presence_penalty=0.0,36 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]
Qwen3.6 does not officially support the soft switch of Qwen3, i.e., /think and /nothink.
Qwen3.6 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}.
Preserve Thinking
By default, only the thinking blocks generated in handling the latest user message is retained, resulting in a pattern commonly as interleaved thinking.
Qwen3.6 has been additionally trained to preserve and leverage thinking traces from historical messages.
You can enable this behavior by setting the preserve_thinking option:
[!Note]
If you are using APIs from Alibaba Cloud Model Studio, in addition to changing model, please use "preserve_thinking": True instead of "chat_template_kwargs": {"preserve_thinking": False}.
This capability is particularly beneficial for agent scenarios, where maintaining full reasoning context can enhance decision consistency and, in many cases, reduce overall token consumption by minimizing redundant reasoning. Additionally, it can improve KV cache utilization, optimizing inference efficiency in both thinking and non-thinking modes.
Agentic Usage
Qwen3.6 excels in tool calling capabilities.
Qwen-Agent
We recommend using Qwen-Agent to quickly build Agent applications with Qwen3.6.
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':'qwen3.6-27b',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':True,18'preserve_thinking':True,19},20},21}2223# Using OpenAI-compatible API endpoint.24# functionality of the deployment frameworks and let Qwen-Agent automate the related operations.25#26# llm_cfg = {27# # Use your own model service compatible with OpenAI API by vLLM/SGLang:28# 'model': 'Qwen/Qwen3.6-27B',29# 'model_type': 'qwenvl_oai',30# 'model_server': 'http://localhost:8000/v1', # api_base31# 'api_key': 'EMPTY',32#33# 'generate_cfg': {34# 'use_raw_api': True,35# # When using vLLM/SGLang OAI API, pass the parameter of whether to enable thinking mode in this way36# 'extra_body': {37# 'chat_template_kwargs': {'enable_thinking': True, 'preserve_thinking': True}38# },39# },40# }4142# Define Tools43tools =[44{'mcpServers':{# You can specify the MCP configuration file45"filesystem":{46"command":"npx",47"args":["-y","@modelcontextprotocol/server-filesystem","/Users/xxxx/Desktop"]48}49}50}51]5253# Define Agent54bot = Assistant(llm=llm_cfg, function_list=tools)5556# Streaming generation57messages =[{'role':'user','content':'Help me organize my desktop.'}]58for responses in bot.run(messages=messages):59pass60print(responses)6162# Streaming generation63messages =[{'role':'user','content':'Develop a dog website and save it on the desktop'}]64for responses in bot.run(messages=messages):65pass66print(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.
Qwen3.6 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=0.0, 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"."
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,
