    ██╗    ██╗██╗  ██╗ █████╗  ██╗ ██████╗ 
    ██║    ██║██║  ██║██╔══██╗███║██╔════╝ 
    ██║ █╗ ██║███████║███████║╚██║███████╗ 
    ██║███╗██║╚════██║██╔══██║ ██║██╔═══██╗
    ╚███╔███╔╝     ██║██║  ██║ ██║╚██████╔╝
     ╚══╝╚══╝      ╚═╝╚═╝  ╚═╝ ╚═╝ ╚═════╝ 
        🗜️ COMPRESSED & OPTIMIZED 🚀

Qwen3-30B-A3B-Thinking-2507 - W4A16 Quantized

W4A16 (4-bit weights, 16-bit activations) quantized version of Qwen/Qwen3-30B-A3B-Thinking-2507 using LLM-Compressor.

🗜️ Memory: ~75% reduction vs FP16
🚀 Speed: Faster inference on supported hardware
🔗 Original model: Qwen/Qwen3-30B-A3B-Thinking-2507

Click to view compression code

from datasets import load_dataset
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor import oneshot
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model with memory management
model_stub = "Qwen/Qwen3-30B-A3B-Thinking-2507"
model_name = model_stub.split("/")[-1]

# Use conservative parameters
num_samples = 1024
max_seq_len = 8192

print(f"Loading model: {model_stub}")
model = AutoModelForCausalLM.from_pretrained(
    model_stub,
    torch_dtype="auto",
    device_map="auto",
    low_cpu_mem_usage=True,
    max_memory={0: "44GB", "cpu": "55GB"},
)

print("Loading tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(model_stub)

print("Loading calibration dataset...")
def preprocess_fn(example):
    return {"text": tokenizer.apply_chat_template(
        example["messages"], 
        add_generation_prompt=False, 
        tokenize=False
    )}

# Load dataset and preprocess
ds = load_dataset("neuralmagic/LLM_compression_calibration", split=f"train[:{num_samples}]")
ds = ds.map(preprocess_fn)
ds = ds.shuffle(seed=42)

# Tokenize the dataset
def tokenize(sample):
    return tokenizer(
        sample["text"],
        padding=False,
        max_length=max_seq_len,
        truncation=True,
        add_special_tokens=False,
    )

print("Tokenizing dataset...")
ds = ds.map(tokenize, remove_columns=ds.column_names)

# Configure GPTQ with proper Qwen3 MoE ignore patterns
print("Configuring quantization recipe...")
recipe = GPTQModifier(
    targets="Linear",
    scheme="W4A16", 
    ignore=["lm_head", "re:.*mlp.gate$"],  # Qwen3 MoE pattern (no shared experts)
    dampening_frac=0.01,
    # Remove sequential_targets - let llmcompressor handle automatically
)

# Apply quantization
print("Starting quantization process...")
oneshot(
    model=model,
    dataset=ds,
    recipe=recipe,
    max_seq_length=max_seq_len,
    num_calibration_samples=num_samples,
)

# Save quantized model
save_path = model_name + "-gptq-w4a16"
print(f"Saving model to: {save_path}")
model.save_pretrained(save_path, save_compressed=True)
tokenizer.save_pretrained(save_path)

print("Quantization completed successfully!")

📄 Original Model README

Qwen3-30B-A3B-Thinking-2507

Highlights

Over the past three months, we have continued to scale the thinking capability of Qwen3-30B-A3B, improving both the quality and depth of reasoning. We are pleased to introduce Qwen3-30B-A3B-Thinking-2507, featuring the following key enhancements:

Significantly improved performance on reasoning tasks, including logical reasoning, mathematics, science, coding, and academic benchmarks that typically require human expertise.
Markedly better general capabilities, such as instruction following, tool usage, text generation, and alignment with human preferences.
Enhanced 256K long-context understanding capabilities.

NOTE: This version has an increased thinking length. We strongly recommend its use in highly complex reasoning tasks.

Model Overview

Qwen3-30B-A3B-Thinking-2507 has the following features:

Type: Causal Language Models
Training Stage: Pretraining & Post-training
Number of Parameters: 30.5B in total and 3.3B activated
Number of Paramaters (Non-Embedding): 29.9B
Number of Layers: 48
Number of Attention Heads (GQA): 32 for Q and 4 for KV
Number of Experts: 128
Number of Activated Experts: 8
Context Length: 262,144 natively.

NOTE: This model supports only thinking mode. Meanwhile, specifying enable_thinking=True is no longer required.

Additionally, to enforce model thinking, the default chat template automatically includes <think>. Therefore, it is normal for the model's output to contain only </think> without an explicit opening <think> tag.

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

Performance

	Gemini2.5-Flash-Thinking	Qwen3-235B-A22B Thinking	Qwen3-30B-A3B Thinking	Qwen3-30B-A3B-Thinking-2507
Knowledge
MMLU-Pro	81.9	82.8	78.5	80.9
MMLU-Redux	92.1	92.7	89.5	91.4
GPQA	82.8	71.1	65.8	73.4
SuperGPQA	57.8	60.7	51.8	56.8
Reasoning
AIME25	72.0	81.5	70.9	85.0
HMMT25	64.2	62.5	49.8	71.4
LiveBench 20241125	74.3	77.1	74.3	76.8
Coding
LiveCodeBench v6 (25.02-25.05)	61.2	55.7	57.4	66.0
CFEval	1995	2056	1940	2044
OJBench	23.5	25.6	20.7	25.1
Alignment
IFEval	89.8	83.4	86.5	88.9
Arena-Hard v2$	56.7	61.5	36.3	56.0
Creative Writing v3	85.0	84.6	79.1	84.4
WritingBench	83.9	80.3	77.0	85.0
Agent
BFCL-v3	68.6	70.8	69.1	72.4
TAU1-Retail	65.2	54.8	61.7	67.8
TAU1-Airline	54.0	26.0	32.0	48.0
TAU2-Retail	66.7	40.4	34.2	58.8
TAU2-Airline	52.0	30.0	36.0	58.0
TAU2-Telecom	31.6	21.9	22.8	26.3
Multilingualism
MultiIF	74.4	71.9	72.2	76.4
MMLU-ProX	80.2	80.0	73.1	76.4
INCLUDE	83.9	78.7	71.9	74.4
PolyMATH	49.8	54.7	46.1	52.6

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

& For highly challenging tasks (including PolyMATH and all reasoning and coding tasks), we use an output length of 81,920 tokens. For all other tasks, we set the output length to 32,768.

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/Qwen3-30B-A3B-Thinking-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=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 

# parsing thinking content
try:
    # rindex finding 151668 (</think>)
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")

print("thinking content:", thinking_content) # no opening <think> tag
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/Qwen3-30B-A3B-Thinking-2507 --context-length 262144  --reasoning-parser deepseek-r1

vLLM:

vllm serve Qwen/Qwen3-30B-A3B-Thinking-2507 --max-model-len 262144 --enable-reasoning --reasoning-parser deepseek_r1

Note: If you encounter out-of-memory (OOM) issues, you may consider reducing the context length to a smaller value. However, since the model may require longer token sequences for reasoning, we strongly recommend using a context length greater than 131,072 when possible.

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
# Using Alibaba Cloud Model Studio
llm_cfg = {
    'model': 'qwen3-30b-a3b-thinking-2507',
    'model_type': 'qwen_dashscope',
}

# Using OpenAI-compatible API endpoint. It is recommended to disable the reasoning and the tool call parsing
# functionality of the deployment frameworks and let Qwen-Agent automate the related operations. For example, 
# `VLLM_USE_MODELSCOPE=true vllm serve Qwen/Qwen3-30B-A3B-Thinking-2507 --served-model-name Qwen3-30B-A3B-Thinking-2507 --tensor-parallel-size 8 --max-model-len 262144`.
#
# llm_cfg = {
#     'model': 'Qwen3-30B-A3B-Thinking-2507',
# 
#     # Use a custom endpoint compatible with OpenAI API:
#     'model_server': 'http://localhost:8000/v1',  # api_base without reasoning and tool call parsing
#     'api_key': 'EMPTY',
#     'generate_cfg': {
#         'thought_in_content': True,
#     },
# }


# 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)

Best Practices

To achieve optimal performance, we recommend the following settings:

Sampling Parameters:
- We suggest using Temperature=0.6, TopP=0.95, 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.
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.

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}, 
}

ramblingpolymath
/

Qwen3-30B-A3B-Thinking-2507-W4A16