vllm serve pahajokiconsulting/Qwen3.6-35B-A3B-MXFP4 \
  --tensor-parallel-size 2 \
  --max-model-len 32768 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_xml \
  --num-speculative-tokens 4 \
  --port 8000

Leave CUDA graphs on (i.e. do not pass --enforce-eager) for best decode throughput. The vision tower is retained in BF16, so --language-model-only is not required.

Speculative decoding (MTP)

The quant is prepared so Qwen3.6's built-in MTP (multi-token prediction) draft model loads correctly: the mtp.* block is kept in BF16 with per-expert .weight tensors, which is the layout stock vLLM's MTP loader expects. Run with --num-speculative-tokens 4 to activate it.

Hardware tested

2× AMD Radeon AI PRO R9700 (64 GB total VRAM), TP=2
Peak model VRAM: 13.38 GiB per rank after MTP draft load

Throughput (single-request decode)

Measured on a 16k-token completion, TP=2, MTP with 4 draft tokens:

Metric	Value
Generation throughput	~153 tokens/s
Mean MTP acceptance length	4.25 / 4
Average draft acceptance	81.2%
Per-position acceptance	93.1% / 85.3% / 75.9% / 70.4%

First request after server start is slower (Triton kernel JIT + TunableOps warmup). Throughput climbs into steady state over a few requests.

For reference, our earlier Qwen3.5-35B-A3B MXFP4 release peaked at 101 t/s after a two-hour dedicated TunableOps tuning pass on the same hardware. The 3.6 number above was observed with no Qwen3.6-specific tuning — TunableOps entries from the 3.5 run carry over because both models share the same GEMM shape profile. Running a dedicated 3.6 tuning pass is likely to push this higher.

Quantization details

Tool: olka/qstream (stock v0.1.0 produces a broken output for Qwen3.6 — a fix to the default exclude list and to 3D-fused passthrough handling is required, see PR linked in that repo)
Input: BF16 (Qwen/Qwen3.6-35B-A3B)
Output: compressed-tensors MXFP4, CT per-expert format
MSE-optimal scale selection enabled
Size: 71.9 GB → 23.1 GB (~32%)

What is MXFP4, what is BF16

The following are kept in BF16 (not quantized) — quantizing any of these breaks loading against stock vLLM:

model.embed_tokens, lm_head
model.language_model.layers.*.self_attn.* and linear_attn.{conv1d,norm}
model.language_model.layers.*.mlp.gate (routing gate)
model.language_model.layers.*.mlp.shared_expert_gate (shape [1, 2048] — not a Linear)
model.language_model.layers.*.mlp.shared_expert.{gate,up,down}_proj (shared expert FFN)
model.visual.* — entire vision tower (attention, MLP, merger, pos_embed, patch_embed)
mtp.* — entire MTP draft block, including per-expert FFN

Quantized to MXFP4 (CT per-expert, .weight_packed + .weight_scale):

model.language_model.layers.*.linear_attn.in_proj_{a,b,qkv,z} and out_proj
model.language_model.layers.*.mlp.experts.{i}.{gate,up,down}_proj — all 256 experts

Why the extended exclude list

Qwen3.6 introduces a shared_expert_gate submodule that Qwen3.5 also has, but whose name doesn't match qstream's default *.mlp.gate. pattern. Attempting to quantize it produces shared_expert_gate.weight_packed / .weight_scale tensors that vLLM's loader silently skips (not found in params_dict), leaving the shared-expert gate uninitialized and producing garbage output. The vision tower and MTP block have analogous issues — vLLM expects those modules to be BF16 dense, not compressed-tensors. The working Qwen3.5 MXFP4 release kept all of these in BF16; this release does the same.

Qwen3.6-35B-A3B

[!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.

For more details, please refer to our blog post Qwen3.6-35B-A3B.

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
- Hidden Dimension: 2048
- Token Embedding: 248320 (Padded)
- Number of Layers: 40
- Hidden Layout: 10 × (3 × (Gated DeltaNet → MoE) → 1 × (Gated Attention → MoE))
- Gated DeltaNet:
  - 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

	Qwen3.5-27B	Gemma4-31B	Qwen3.5-35BA3B	Gemma4-26BA4B	Qwen3.6-35BA3B
Coding Agent
SWE-bench Verified	75.0	52.0	70.0	17.4	73.4
SWE-bench Multilingual	69.3	51.7	60.3	17.3	67.2
SWE-bench Pro	51.2	35.7	44.6	13.8	49.5
Terminal-Bench 2.0	41.6	42.9	40.5	34.2	51.5
Claw-Eval _Avg	64.3	48.5	65.4	58.8	68.7
Claw-Eval _Pass^3	46.2	25.0	51.0	28.0	50.0
SkillsBench _Avg5	27.2	23.6	4.4	12.3	28.7
QwenClawBench	52.2	41.7	47.7	38.7	52.6
NL2Repo	27.3	15.5	20.5	11.6	29.4
QwenWebBench	1068	1197	978	1178	1397
General Agent
TAU3-Bench	68.4	67.5	68.9	59.0	67.2
VITA-Bench	41.8	43.0	29.1	36.9	35.6
DeepPlanning	22.6	24.0	22.8	16.2	25.9
Tool Decathlon	31.5	21.2	28.7	12.0	26.9
MCPMark	36.3	18.1	27.0	14.2	37.0
MCP-Atlas	68.4	57.2	62.4	50.0	62.8
WideSearch	66.4	35.2	59.1	38.3	60.1
Knowledge
MMLU-Pro	86.1	85.2	85.3	82.6	85.2
MMLU-Redux	93.2	93.7	93.3	92.7	93.3
SuperGPQA	65.6	65.7	63.4	61.4	64.7
C-Eval	90.5	82.6	90.2	82.5	90.0
STEM & Reasoning
GPQA	85.5	84.3	84.2	82.3	86.0
HLE	24.3	19.5	22.4	8.7	21.4
LiveCodeBench v6	80.7	80.0	74.6	77.1	80.4
HMMT Feb 25	92.0	88.7	89.0	91.7	90.7
HMMT Nov 25	89.8	87.5	89.2	87.5	89.1
HMMT Feb 26	84.3	77.2	78.7	79.0	83.6
IMOAnswerBench	79.9	74.5	76.8	74.3	78.9
AIME26	92.6	89.2	91.0	88.3	92.7

* 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: An internal real-user-distribution Claw agent benchmark (open-sourcing soon); 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.
* TAU3-Bench: We use the official user model (gpt-5.2, low reasoning effort) + default BM25 retrieval.
* VITA-Bench: Avg subdomain scores; using claude-4-sonnet as judger, as the official judger (claude-3.7-sonnet) is no longer available.
* MCPMark: GitHub MCP v0.30.3; Playwright responses truncated at 32K tokens.
* MCP-Atlas: Public set score; gemini-2.5-pro judger.
* 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	Claude-Sonnet-4.5	Gemma4-31B	Gemma4-26BA4B	Qwen3.5-35B-A3B	Qwen3.6-35B-A3B
STEM and Puzzle
MMMU	82.3	79.6	80.4	78.4	81.4	81.7
MMMU-Pro	75.0	68.4	76.9*	73.8*	75.1	75.3
Mathvista(mini)	87.8	79.8	79.3	79.4	86.2	86.4
ZEROBench_sub	36.2	26.3	26.0	26.3	34.1	34.4
General VQA
RealWorldQA	83.7	70.3	72.3	72.2	84.1	85.3
MMBench_EN-DEV-v1.1	92.6	88.3	90.9	89.0	91.5	92.8
SimpleVQA	56.0	57.6	52.9	52.2	58.3	58.9
HallusionBench	70.0	59.9	67.4	66.1	67.9	69.8
Text Recognition and Document Understanding
OmniDocBench1.5	88.9	85.8	80.1	74.4	89.3	89.9
CharXiv(RQ)	79.5	67.2	67.9	69.0	77.5	78.0
CC-OCR	81.0	68.1	75.7	74.5	80.7	81.9
AI2D_TEST	92.9	87.0	89.0	88.3	92.6	92.7
Spatial Intelligence
RefCOCO(avg)	90.9	--	--	--	89.2	92.0
ODInW13	41.1	--	--	--	42.6	50.8
EmbSpatialBench	84.5	71.8	--	--	83.1	84.3
RefSpatialBench	67.7	--	--	--	63.5	64.3
Video Understanding
VideoMME_{(w sub.)}	87.0	81.1	--	--	86.6	86.6
VideoMME_{(w/o sub.)}	82.8	75.3	--	--	82.5	82.5
VideoMMMU	82.3	77.6	81.6	76.0	80.4	83.7
MLVU	85.9	72.8	--	--	85.6	86.2
MVBench	74.6	--	--	--	74.8	74.6
LVBench	73.6	--	--	--	71.4	71.4

* Empty cells (--) indicate scores not 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:

uv pip install 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/Qwen3.6-35B-A3B --port 8000 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3
Tool Use: To support tool use, you can use the following command.

python -m sglang.launch_server --model-path Qwen/Qwen3.6-35B-A3B --port 8000 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --tool-call-parser qwen3_coder
Multi-Token Prediction (MTP): The following command is recommended for MTP:

python -m sglang.launch_server --model-path Qwen/Qwen3.6-35B-A3B --port 8000 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --speculative-algo NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4

For detailed deployment guide, see the SGLang Qwen3.5 Cookbook.

vLLM

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:

uv pip install vllm --torch-backend=auto

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.

vllm serve Qwen/Qwen3.6-35B-A3B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3
Tool Call: To support tool use, you can use the following command.

vllm serve Qwen/Qwen3.6-35B-A3B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
Multi-Token Prediction (MTP): The following command is recommended for MTP:

vllm serve Qwen/Qwen3.6-35B-A3B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --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/Qwen3.6-35B-A3B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --language-model-only

For detailed deployment guide, see the vLLM Qwen3.5 Recipe.

KTransformers

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:

transformers serve Qwen/Qwen3.6-35B-A3B --port 8000 --continuous-batching

Using Qwen3.6 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.:

shell

pip install -U openai

# Set the following accordingly
export OPENAI_BASE_URL="http://localhost:8000/v1"
export OPENAI_API_KEY="EMPTY"

[!Tip] We recommend using the following set of sampling parameters for generation

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

Thinking mode for precise coding tasks (e.g. WebDev): temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0

Instruct (or non-thinking) mode for general tasks: temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0

Instruct (or non-thinking) mode for reasoning tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0

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.

Text-Only Input

python

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {"role": "user", "content": "Type \"I love Qwen3.6\" backwards"},
]

chat_response = client.chat.completions.create(
    model="Qwen/Qwen3.6-35B-A3B",
    messages=messages,
    max_tokens=81920,
    temperature=1.0,
    top_p=0.95,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
    }, 
)
print("Chat response:", chat_response)

Image Input

python

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/CI_Demo/mathv-1327.jpg"
                }
            },
            {
                "type": "text",
                "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) $1+\\sqrt{2}$"
            }
        ]
    }
]

response = client.chat.completions.create(
    model="Qwen/Qwen3.6-35B-A3B",
    messages=messages,
    max_tokens=81920,
    temperature=1.0,
    top_p=0.95,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
    }, 
)
print("Chat response:", chat_response)

Video Input

python

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": "How many porcelain jars were discovered in the niches located in the primary chamber of the tomb?"
            }
        ]
    }
]

# 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.
response = client.chat.completions.create(
    model="Qwen/Qwen3.6-35B-A3B",
    messages=messages,
    max_tokens=81920,
    temperature=1.0,
    top_p=0.95,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
        "mm_processor_kwargs": {"fps": 2, "do_sample_frames": True},
    }, 
)

print("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,

python

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/Qwen3.6/demo/RealWorld/RealWorld-04.png"
                }
            },
            {
                "type": "text",
                "text": "Where is this?"
            }
        ]
    }
]

chat_response = client.chat.completions.create(
    model="Qwen/Qwen3.6-35B-A3B",
    messages=messages,
    max_tokens=32768,
    temperature=0.7,
    top_p=0.8,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
        "chat_template_kwargs": {"enable_thinking": False},
    }, 
)
print("Chat response:", chat_response)

[!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:

python

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [...]

chat_response = client.chat.completions.create(
    model="Qwen/Qwen3.6-35B-A3B",
    messages=messages,
    max_tokens=32768,
    temperature=0.7,
    top_p=0.8,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
        "chat_template_kwargs": {"preserve_thinking": True},
    }, 
)
print("Chat response:", chat_response)

[!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

import os
from qwen_agent.agents import Assistant

# Define LLM
# Using Alibaba Cloud Model Studio
llm_cfg = {
    # Use the OpenAI-compatible model service provided by DashScope:
    'model': 'Qwen3.6-35B-A3B',
    'model_type': 'qwenvl_oai',
    'model_server': 'https://dashscope.aliyuncs.com/compatible-mode/v1',
    'api_key': os.getenv('DASHSCOPE_API_KEY'),

    'generate_cfg': {
        'use_raw_api': True,
        # When using Dash Scope OAI API, pass the parameter of whether to enable thinking mode in this way
        'extra_body': {
            'enable_thinking': True,
            'preserve_thinking': True,
        },
    },
}

# Using OpenAI-compatible API endpoint.
# functionality of the deployment frameworks and let Qwen-Agent automate the related operations.
#
# llm_cfg = {
#     # Use your own model service compatible with OpenAI API by vLLM/SGLang:
#     'model': 'Qwen/Qwen3.6-35B-A3B',
#     'model_type': 'qwenvl_oai',
#     'model_server': 'http://localhost:8000/v1',  # api_base
#     'api_key': 'EMPTY',
#
#     'generate_cfg': {
#         'use_raw_api': True,
#         # When using vLLM/SGLang OAI API, pass the parameter of whether to enable thinking mode in this way
#         'extra_body': {
#             'chat_template_kwargs': {'enable_thinking': True, 'preserve_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.

Processing Ultra-Long Texts

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:

json
1{ 2 "mrope_interleaved": true, 3 "mrope_section": [ 4 11, 5 11, 6 10 7 ], 8 "rope_type": "yarn", 9 "rope_theta": 10000000, 10 "partial_rotary_factor": 0.25, 11 "factor": 4.0, 12 "original_max_position_embeddings": 262144, 13}
Passing command line arguments:

For vllm, you can use

VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve ... --hf-overrides '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --max-model-len 1010000

For sglang and ktransformers, you can use

SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 python -m sglang.launch_server ... --json-model-override-args '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --context-length 1010000

[!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
  - Thinking mode for precise coding tasks (e.g., WebDev):
    temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0
  - Instruct (or non-thinking) mode for general tasks:
    temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
  - Instruct (or non-thinking) mode for reasoning tasks:
    temperature=1.0, top_p=1.0, top_k=40, min_p=0.0, presence_penalty=2.0, 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,

{"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.

bibtex

@misc{qwen36_35b_a3b,
    title = {{Qwen3.6-35B-A3B}: Agentic Coding Power, Now Open to All},
    url = {https://qwen.ai/blog?id=qwen3.6-35b-a3b},
    author = {{Qwen Team}},
    month = {April},
    year = {2026}
}