Instructions to use aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4")
model = AutoModelForCausalLM.from_pretrained("aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4

SGLang

How to use aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4 with Docker Model Runner:
```
docker model run hf.co/aws-prototyping/Qwen3-235B-A22B-Instruct-2507_MXFP4
```

Qwen3-235B-A22B-Instruct-2507_MXFP4

This checkpoint is a variant of Qwen3-235B-A22B-Instruct-2507, where expert weights have been quantized to MXFP4 format similarly to gpt-oss-20b and gpt-oss-120b.

For quantizing weights we used the function downcast_to_mxfp from triton-kernels.

The checkpoint might come with a small drop in accuracy, but has ~71% size reduction compared to the original BF16 checkpoint.

Accuracy Comparison

Model	GSM8K (strict-match)	GSM8K (flexible-extract)
Qwen3-235B-A22B-Instruct-2507 (BF16)	90.14% ± 0.82%	91.05% ± 0.79%
Qwen3-235B-A22B-Instruct-2507_MXFP4	90.45% ± 0.81%	91.36% ± 0.77%

Checkpoint Size

Model	Size	Reduction
Qwen3-235B-A22B-Instruct-2507 (BF16)	438 GB	-
Qwen3-235B-A22B-Instruct-2507_MXFP4	128 GB	71% smaller

Downloads last month: 4

Safetensors

Model size

242B params

Tensor type

BF16