Qwen3-235B-A22B-Instruct-2507-INT8

Qwen3-235B-A22B-Instruct-2507-INT8

W8A8 INT8 Quantized (using GPTQ quantization) version of Qwen/Qwen3-235B-A22B-Instruct-2507

Made for inference using vLLM on GPU without FP8 support.

Base model Qwen/Qwen3-235B-A22B-Instruct-2507

Qwen3-235B-A22B-Instruct-2507

<a href="https://chat.qwen.ai/" target="_blank" style="margin: 2px;"> <img alt="Chat" src="https://img.shields.io/badge/%F0%9F%92%9C%EF%B8%8F%20Qwen%20Chat%20-536af5" style="display: inline-block; vertical-align: middle;"/> </a>

Highlights

We introduce the updated version of the Qwen3-235B-A22B non-thinking mode, named Qwen3-235B-A22B-Instruct-2507, featuring the following key enhancements:

• Significant improvements in general capabilities, including instruction following, logical reasoning, text comprehension, mathematics, science, coding and tool usage.
• Substantial gains in long-tail knowledge coverage across multiple languages.
• Markedly better alignment with user preferences in subjective and open-ended tasks, enabling more helpful responses and higher-quality text generation.
• Enhanced capabilities in 256K long-context understanding.

Model Overview

Qwen3-235B-A22B-Instruct-2507 has the following features:

• Type: Causal Language Models
• Training Stage: Pretraining & Post-training
• Number of Parameters: 235B in total and 22B activated
• Number of Paramaters (Non-Embedding): 234B
• Number of Layers: 94
• Number of Attention Heads (GQA): 64 for Q and 4 for KV
• Number of Experts: 128
• Number of Activated Experts: 8
• Context Length: 262,144 natively and extendable up to 1,010,000 tokens

NOTE: This model supports only non-thinking mode and does not generate <think></think> blocks in its output. Meanwhile, specifying enable_thinking=False is no longer required.

For more details, including benchmark evaluation, hardware requirements, and inference performance, please refer to our blog, GitHub, and Documentation.

Performance

*: For reproducibility, we report the win rates evaluated by GPT-4.1.

#: Results were generated using GPT-4o-20241120, as access to the native function calling API of GPT-4o-0327 was unavailable.

Quickstart

The code of Qwen3-MoE has been in the latest Hugging Face transformers and we advise you to use the latest version of transformers.

With transformers<4.51.0, you will encounter the following error:

KeyError: 'qwen3_moe'

The following contains a code snippet illustrating how to use the model generate content based on given inputs.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Qwen/Qwen3-235B-A22B-Instruct-2507"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# prepare the model input
prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=16384
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 

content = tokenizer.decode(output_ids, skip_special_tokens=True)

print("content:", content)

For deployment, you can use sglang>=0.4.6.post1 or vllm>=0.8.5 or to create an OpenAI-compatible API endpoint:

• SGLang:

  python -m sglang.launch_server --model-path Qwen/Qwen3-235B-A22B-Instruct-2507 --tp 8 --context-length 262144
  

• vLLM:

  vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507 --tensor-parallel-size 8 --max-model-len 262144
  

Note: If you encounter out-of-memory (OOM) issues, consider reducing the context length to a shorter value, such as 32,768.

For local use, applications such as Ollama, LMStudio, MLX-LM, llama.cpp, and KTransformers have also supported Qwen3.

Agentic Use

Qwen3 excels in tool calling capabilities. We recommend using Qwen-Agent to make the best use of agentic ability of Qwen3. Qwen-Agent encapsulates tool-calling templates and tool-calling parsers internally, greatly reducing coding complexity.

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.

from qwen_agent.agents import Assistant

# Define LLM
llm_cfg = {
    'model': 'Qwen3-235B-A22B-Instruct-2507',

    # Use a custom endpoint compatible with OpenAI API:
    'model_server': 'http://localhost:8000/v1',  # api_base
    'api_key': 'EMPTY',
}

# Define Tools
tools = [
    {'mcpServers': {  # You can specify the MCP configuration file
            'time': {
                'command': 'uvx',
                'args': ['mcp-server-time', '--local-timezone=Asia/Shanghai']
            },
            "fetch": {
                "command": "uvx",
                "args": ["mcp-server-fetch"]
            }
        }
    },
  'code_interpreter',  # Built-in tools
]

# Define Agent
bot = Assistant(llm=llm_cfg, function_list=tools)

# Streaming generation
messages = [{'role': 'user', 'content': 'https://qwenlm.github.io/blog/ Introduce the latest developments of Qwen'}]
for responses in bot.run(messages=messages):
    pass
print(responses)

Processing Ultra-Long Texts

To support ultra-long context processing (up to 1 million tokens), we integrate two key techniques:

• Dual Chunk Attention (DCA): A length extrapolation method that splits long sequences into manageable chunks while preserving global coherence.
• MInference: A sparse attention mechanism that reduces computational overhead by focusing on critical token interactions.

Together, these innovations significantly improve both generation quality and inference efficiency for sequences beyond 256K tokens. On sequences approaching 1M tokens, the system achieves up to a 3× speedup compared to standard attention implementations.

For full technical details, see the Qwen2.5-1M Technical Report.

How to Enable 1M Token Context

[!NOTE] To effectively process a 1 million token context, users will require approximately 1000 GB of total GPU memory. This accounts for model weights, KV-cache storage, and peak activation memory demands.

Step 1: Update Configuration File

Download the model and replace the content of your config.json with config_1m.json, which includes the config for length extrapolation and sparse attention.

export MODELNAME=Qwen3-235B-A22B-Instruct-2507
huggingface-cli download Qwen/${MODELNAME} --local-dir ${MODELNAME}
mv ${MODELNAME}/config.json ${MODELNAME}/config.json.bak
mv ${MODELNAME}/config_1m.json ${MODELNAME}/config.json

Step 2: Launch Model Server

After updating the config, proceed with either vLLM or SGLang for serving the model.

Option 1: Using vLLM

To run Qwen with 1M context support:

pip install -U vllm \
    --torch-backend=auto \
    --extra-index-url https://wheels.vllm.ai/nightly

Then launch the server with Dual Chunk Flash Attention enabled:

VLLM_ATTENTION_BACKEND=DUAL_CHUNK_FLASH_ATTN VLLM_USE_V1=0 \
vllm serve ./Qwen3-235B-A22B-Instruct-2507 \
  --tensor-parallel-size 8 \
  --max-model-len 1010000 \
  --enable-chunked-prefill \
  --max-num-batched-tokens 131072 \
  --enforce-eager \
  --max-num-seqs 1 \
  --gpu-memory-utilization 0.85

Key Parameters

Option 2: Using SGLang

First, clone and install the specialized branch:

git clone https://github.com/sgl-project/sglang.git
cd sglang
pip install -e "python[all]"

Launch the server with DCA support:

python3 -m sglang.launch_server \
    --model-path ./Qwen3-235B-A22B-Instruct-2507 \
    --context-length 1010000 \
    --mem-frac 0.75 \
    --attention-backend dual_chunk_flash_attn \
    --tp 8 \
    --chunked-prefill-size 131072

Key Parameters

Troubleshooting:

1. Encountering the error: "The model's max sequence length (xxxxx) is larger than the maximum number of tokens that can be stored in the KV cache." or "RuntimeError: Not enough memory. Please try to increase --mem-fraction-static."

   The VRAM reserved for the KV cache is insufficient.

   • vLLM: Consider reducing the max_model_len or increasing the tensor_parallel_size and gpu_memory_utilization. Alternatively, you can reduce max_num_batched_tokens, although this may significantly slow down inference.
   • SGLang: Consider reducing the context-length or increasing the tp and mem-frac. Alternatively, you can reduce chunked-prefill-size, although this may significantly slow down inference.

2. Encountering the error: "torch.OutOfMemoryError: CUDA out of memory."

   The VRAM reserved for activation weights is insufficient. You can try lowering gpu_memory_utilization or mem-frac, but be aware that this might reduce the VRAM available for the KV cache.

3. Encountering the error: "Input prompt (xxxxx tokens) + lookahead slots (0) is too long and exceeds the capacity of the block manager." or "The input (xxx xtokens) is longer than the model's context length (xxx tokens)."

   The input is too lengthy. Consider using a shorter sequence or increasing the max_model_len or context-length.

Long-Context Performance

We test the model on an 1M version of the RULER benchmark.

• All models are evaluated with Dual Chunk Attention enabled.
• Since the evaluation is time-consuming, we use 260 samples for each length (13 sub-tasks, 20 samples for each).

Best Practices

To achieve optimal performance, we recommend the following settings:

1. Sampling Parameters:

   • We suggest using Temperature=0.7, TopP=0.8, TopK=20, and MinP=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.

2. Adequate Output Length: We recommend using an output length of 16,384 tokens for most queries, which is adequate for instruct models.

3. 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"."

Citation

If you find our work helpful, feel free to give us a cite.

@misc{qwen3technicalreport,
      title={Qwen3 Technical Report}, 
      author={Qwen Team},
      year={2025},
      eprint={2505.09388},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.09388}, 
}

@article{qwen2.5-1m,
      title={Qwen2.5-1M Technical Report}, 
      author={An Yang and Bowen Yu and Chengyuan Li and Dayiheng Liu and Fei Huang and Haoyan Huang and Jiandong Jiang and Jianhong Tu and Jianwei Zhang and Jingren Zhou and Junyang Lin and Kai Dang and Kexin Yang and Le Yu and Mei Li and Minmin Sun and Qin Zhu and Rui Men and Tao He and Weijia Xu and Wenbiao Yin and Wenyuan Yu and Xiafei Qiu and Xingzhang Ren and Xinlong Yang and Yong Li and Zhiying Xu and Zipeng Zhang},
      journal={arXiv preprint arXiv:2501.15383},
      year={2025}
}