gpt-oss-20b

This repository contains openai/gpt-oss-20b together with a Furiosa Executable Bundle (FXB) for running it on FuriosaAI RNGD with Furiosa-LLM. The same model also runs on other frameworks (such as vLLM, SGLang, and Transformers); for usage with those, see the upstream openai/gpt-oss-20b model card.

Overview

gpt-oss-20b is OpenAI's smaller open-weight reasoning model, designed for lower-latency and on-device use cases. It is an auto-regressive Mixture-of-Experts (MoE) transformer that emits responses in the harmony response format, separating the model's chain-of-thought reasoning from its final answer and carrying native support for tool calling and configurable reasoning effort. Its intended use is the same as the upstream openai/gpt-oss-20b, and it is released under the Apache 2.0 License.

• Architecture: GPT-OSS (Mixture-of-Experts)
• Input / Output: Text / Text
• Supported Inference Engine: Furiosa LLM
• Supported Hardware: FuriosaAI RNGD

Quantization

The MoE expert weights are quantized to MXFP4, the format gpt-oss ships in upstream. The remaining components (attention, router, and embeddings) stay in higher precision.

Features

• Reasoning. gpt-oss-20b is a reasoning model that first produces a chain of thought and then a final answer. Because it uses the harmony response format, Furiosa-LLM detects it automatically and parses the reasoning out of the box — no --reasoning-parser flag is required. The reasoning depth is controllable per request through reasoning_effort ("low", "medium", "high").
• Tool calling. The model supports tool (function) calling through the openai tool-call parser

Parallelism Strategy

On RNGD, gpt-oss-20b runs with a tensor-parallel size of 8 PEs, which maps to a single RNGD card (8 PEs per card).

Usage

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

Launch the server

The simplest way to serve the model is:

# Launch the server, listening on port 8000 by default
furiosa-llm serve furiosa-ai/gpt-oss-20b

Reasoning works out of the box (the harmony format is auto-detected), so no --reasoning-parser flag is needed. The reasoning content is returned in a separate field (see Basic Usage below).

To also enable tool (function) calling, add the openai tool-call parser:

furiosa-llm serve furiosa-ai/gpt-oss-20b \
  --enable-auto-tool-choice \
  --tool-call-parser openai

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:

curl http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
    "model": "furiosa-ai/gpt-oss-20b",
    "messages": [{"role": "user", "content": "What is the capital of France?"}]
    }' \
    | python -m json.tool

gpt-oss-20b returns its reasoning 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/gpt-oss-20b",
    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 OpenAI recommends (also used by the OpenAI Agents SDK, vLLM, and others). It appears only in responses that contain reasoning content; accessing it otherwise raises an AttributeError.

Advanced Usage

Reasoning effort. You can control how much effort the model spends reasoning with the reasoning_effort parameter ("low", "medium", or "high"):

# Request high reasoning effort
response = client.chat.completions.create(
    model="furiosa-ai/gpt-oss-20b",
    messages=[{"role": "user", "content": "How many r's are in 'strawberry'?"}],
    extra_body={"reasoning_effort": "high"},
)
print(response.choices[0].message.content)

Tool calling. With the server launched using --enable-auto-tool-choice --tool-call-parser openai (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

• Tool Calling — parsers, tool-choice options, and more examples
• Furiosa-LLM Server (furiosa-llm serve) — full OpenAI-compatible API reference and serving options
• openai/gpt-oss-20b — upstream model card