Instructions to use Qwen/Qwen3.8-2.4T-A95B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Qwen/Qwen3.8-2.4T-A95B with Transformers:

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

pipe = pipeline("text-generation", model="Qwen/Qwen3.8-2.4T-A95B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.8-2.4T-A95B")
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.8-2.4T-A95B", device_map="auto")
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]:]))

Inference
HuggingChat
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Qwen/Qwen3.8-2.4T-A95B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Qwen/Qwen3.8-2.4T-A95B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Qwen/Qwen3.8-2.4T-A95B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Qwen/Qwen3.8-2.4T-A95B

SGLang

How to use Qwen/Qwen3.8-2.4T-A95B 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 "Qwen/Qwen3.8-2.4T-A95B" \
    --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": "Qwen/Qwen3.8-2.4T-A95B",
		"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 "Qwen/Qwen3.8-2.4T-A95B" \
        --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": "Qwen/Qwen3.8-2.4T-A95B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Qwen/Qwen3.8-2.4T-A95B with Docker Model Runner:
```
docker model run hf.co/Qwen/Qwen3.8-2.4T-A95B
```

Qwen3.8-A50B: Can We Distill A95B into a 50B MoE with 5–12B Active Parameters?

#32

by 88AtlasClaw - opened 22 days ago

Discussion

88AtlasClaw

22 days ago

Is anyone interested in creating a Qwen3.8-based MoE model around 50B total parameters, with only ~5–12B active parameters per token?

The idea would be to use Qwen3.8-A95B as the teacher, applying expert pruning, distillation and/or other compression techniques to preserve as much of its reasoning, coding and agentic capability as possible while producing a model that is actually practical to run locally.

Something like:

Qwen3.8-A95B (2.4T total / 95B active)
↓ expert pruning + distillation
~50B total / ~5–12B active

A model in this range would be extremely interesting for local users with 32–64 GB RAM and modest GPUs (e.g. 6–12 GB VRAM). It could potentially offer a much better quality/compute ratio than a conventional dense 50B model.

I know this would not be a trivial conversion or simple quantization — the goal would be to create a genuinely smaller MoE while retaining as much of the A95B capability as possible.

Would anyone be interested in experimenting with this? I think a Qwen3.8-A50B-ish model with ~5–12B active parameters could be a very interesting community project.

Banaxi-Tech

22 days ago

I would like that, but if then also A18B with 4B active because i have 16G VRAM

CoderX1997

14 days ago

hey... theres a better way to make it smaller...

Try REAP - removing unused / insignificant experts.

Also, if you want to do this, consider that distilling 2.4T paramaters requires a load of GPUs...

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment