Available on request
gpt-oss API:
OpenAI’s open-weight models, on request.
gpt-oss-120b and gpt-oss-20b are OpenAI’s open-weight reasoning models, released under Apache 2.0. If your team already builds on the OpenAI SDK and wants an open-weight model behind the same kind of API, we can talk about adding gpt-oss to the Xinference Model API.
Two sizes
gpt-oss-120b vs gpt-oss-20b
| Detail | gpt-oss-120b | gpt-oss-20b |
|---|---|---|
| Parameters | About 117B | About 21B |
| Context window | 131,072 tokens Model maker’s figure | 131,072 tokens Model maker’s figure |
| Licence | Apache 2.0 | Apache 2.0 |
| Released | August 2025 | August 2025 |
| Where it fits | Harder reasoning and agent steps where accuracy matters more than cost | High-volume, lower-latency calls such as classification, routing and extraction |
Sources: OpenAI’s gpt-oss-120b and gpt-oss-20b model cards, checked 23 October 2026. Parameter counts are rounded from the published weights. gpt-oss, OpenAI and the OpenAI logo are trademarks of OpenAI, shown only to identify the model. Their use does not imply endorsement.
Coming from the OpenAI API
What changes, and what stays the same
The Model API accepts OpenAI-format requests, so most of the work is a configuration change. A few things are worth testing first.
client = OpenAI(
base_url=os.environ["XINFERENCE_BASE_URL"],
api_key=os.environ["XINFERENCE_API_KEY"],
)
response = client.chat.completions.create(
model="gpt-oss-120b", # name confirmed when added
messages=[{"role": "user", "content": "Route this ticket."}],
)| Test before you switch | Why |
|---|---|
| Reasoning output | gpt-oss reasons before it answers. Check how your code handles reasoning text and token counts. |
| Tool calls | Run your own tool schemas end to end, not just a hello-world prompt. |
| Your prompts | Prompts tuned for a closed model can need light rework on an open-weight one. |
More detail in What is an OpenAI-compatible API?
gpt-oss questions
Is gpt-oss the same as ChatGPT or GPT-5?
No. gpt-oss is a separate family of open-weight models that OpenAI published for anyone to download and run. ChatGPT and the models in OpenAI’s API are closed: their weights are not published.
Can I use gpt-oss commercially?
Yes. Both gpt-oss-120b and gpt-oss-20b are released under the Apache 2.0 licence, which permits commercial use, modification and redistribution. OpenAI also publishes a gpt-oss usage policy, so read it alongside the licence.
Is gpt-oss on the Xinference Model API now?
It is available on request. The Model API’s standard models are Kimi K3 and DeepSeek V4.1 Flash. For gpt-oss, tell us which size you need and your expected monthly volume, and we’ll talk through adding it.
Should I pick 120b or 20b?
Start from the task. Multi-step reasoning and agent work usually justify 120b. Short, repetitive calls at high volume usually suit 20b. Some teams run both and route each request by task.
Do you train on my data?
No. We don’t use your prompts or outputs to train or fine-tune models unless you agree in writing. See the Terms of Service.
Last updated