Qwen3 (dense)

Qwen3 (dense)#

Qwen3 is Alibaba’s latest generation of dense, auto-regressive transformer language models with grouped-query attention. Their hallmark is seamless switching between a thinking mode — emitting a chain of thought before the final answer for complex reasoning, math, and coding — and a non-thinking mode for efficient general dialogue, within a single model. They also offer strong tool-calling and agent capabilities and multilingual support.

Furiosa-LLM runs the Qwen3 dense models in FP8 (static FP8 weights with dynamic FP8 activation quantization; the KV cache stays in 16-bit precision). FuriosaAI publishes pre-compiled FP8 builds under the furiosa-ai organization on the Hugging Face Hub, each shipping a Furiosa Executable Bundle (FXB) for running it on FuriosaAI RNGD with Furiosa-LLM. The same upstream weights also run on other frameworks (such as vLLM, SGLang, and Transformers); for usage with those, see the upstream model cards linked below.

This page covers the dense Qwen3 chat models. For the Mixture-of-Experts variants see Qwen3-MoE, and for the embedding, reranking, and vision-language members of the family see Qwen3-Embedding, Qwen3-Reranker, and Qwen3-VL.

Available Models#

Model

Quantization

RNGD cards

Notes

furiosa-ai/Qwen3-4B-FP8

FP8

1

4B params; hybrid thinking / non-thinking

furiosa-ai/Qwen3-8B-FP8

FP8

1

8.2B params; hybrid thinking / non-thinking

furiosa-ai/Qwen3-32B-FP8

FP8

4

32B params; hybrid thinking / non-thinking

  • Architecture: Qwen3 (dense), Qwen3ForCausalLM

  • Input / Output: Text / Text

  • Quantization: Weights are quantized to FP8 (static, fine-grained, block size 128), and activations use dynamic FP8 quantization at runtime (per-token / per-block). The KV cache stays in 16-bit precision.

Usage#

To run these models with Furiosa-LLM, follow the example commands below after installing Furiosa-LLM and its prerequisites.

Launch the server#

Pass the model’s furiosa-ai/<repo> identifier. The examples below use the hybrid Qwen3-8B-FP8 (single RNGD card), which reasons by default and can switch thinking on and off (see Advanced Usage); the larger Qwen3-32B-FP8 runs on four RNGD cards and launches the same way.

Serve it with the qwen3 reasoning parser so the chain of thought is returned in a separate field:

furiosa-llm serve furiosa-ai/Qwen3-8B-FP8 \
  --reasoning-parser qwen3

To also enable tool (function) calling, add the hermes tool-call parser (the parser used by the Qwen3 series):

furiosa-llm serve furiosa-ai/Qwen3-8B-FP8 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser hermes

When the server is ready, you will see:

INFO:     Started server process [27507]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:8000 (Press CTRL+C to quit)

Basic Usage#

The server exposes an OpenAI-compatible API. You can send a request with curl (replace the model id with the variant you launched):

curl http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
    "model": "furiosa-ai/Qwen3-8B-FP8",
    "messages": [{"role": "user", "content": "What is the capital of France?"}]
    }' \
    | python -m json.tool

With --reasoning-parser qwen3, the thinking content is returned separately from the final answer:

  • response.choices[].message.reasoning (non-streaming)

  • response.choices[].delta.reasoning (streaming)

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="furiosa-ai/Qwen3-8B-FP8",
    messages=[{"role": "user", "content": "How many r's are in 'strawberry'?"}],
)

print("Reasoning:", response.choices[0].message.reasoning)
print("Answer:", response.choices[0].message.content)

Note: The reasoning field is not part of the OpenAI API specification but is a widely followed convention (the OpenAI Agents SDK, vLLM, and others). It appears only in responses that contain reasoning content; accessing it otherwise raises an AttributeError.

Advanced Usage#

Toggling thinking. Qwen3 dense models are hybrid: they reason by default and can switch thinking on and off. To turn thinking off for a single request, pass enable_thinking through chat_template_kwargs; the response then carries no reasoning content, so read only message.content:

# Disable thinking for a single request
response = client.chat.completions.create(
    model="furiosa-ai/Qwen3-8B-FP8",
    messages=[{"role": "user", "content": "What is the capital of France?"}],
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)
print(response.choices[0].message.content)

To default every request to non-thinking, launch the server with --default-chat-template-kwargs (a request can still re-enable thinking with its own chat_template_kwargs):

furiosa-llm serve furiosa-ai/Qwen3-8B-FP8 \
  --reasoning-parser qwen3 \
  --default-chat-template-kwargs '{"enable_thinking": false}'

Tool calling. With the server launched using --enable-auto-tool-choice --tool-call-parser hermes (see Launch the server), pass tools in the request and let the model decide when to call them. See the Tool Calling guide for a complete client example and details on tool-choice options.

Learn more#