gpt-oss-20b(free) API

Technical specifications of gpt-oss-20b

What is gpt-oss-20b?

gpt-oss-20b is OpenAI's smaller open-weight reasoning model in the gpt-oss family. CometAPI also offers free use of gpt-oss-20b; the ID is gpt-oss-20b-free. It has 21 billion total parameters but activates about 3.6 billion parameters per forward pass, using a Mixture-of-Experts architecture to reduce inference requirements while retaining substantial reasoning capacity. OpenAI positions it for lower-latency, local, and specialized workloads rather than as a direct replacement for its largest proprietary models.

The model is text-only and is designed for reasoning and agentic workflows. Its open weights can be downloaded, customized, fine-tuned, and deployed on infrastructure controlled by the developer. OpenAI and Hugging Face document native MXFP4 quantization that allows the model to run within about 16 GB of memory, making it unusually accessible for a model in this parameter class.

Main features of gpt-oss-20b

• Efficient MoE architecture: 21B total parameters with only 3.6B active parameters per forward pass, targeting lower latency and lower deployment requirements.
• Configurable reasoning: Developers can select low, medium, or high reasoning effort to trade response latency and compute for reasoning quality.
• Agentic tool use: The model supports workflows involving function calling, web browsing, Python execution, and structured outputs when the serving runtime exposes those tools.
• Open-weight customization: The Apache 2.0 license permits broad modification and commercial deployment, subject to the gpt-oss usage policy.
• Local and private deployment: The model is intended to run on infrastructure controlled by developers, including local or private environments.
• Long context: The production model documentation lists a 131,072-token context window and 131,072 maximum output-token limit.

Benchmark performance of gpt-oss-20b

OpenAI's published evaluation results show that gpt-oss-20b is particularly strong on mathematical reasoning and coding relative to its size. At high reasoning effort with tools, it reaches 98.7% on AIME 2025, 96.0% on AIME 2024, 60.7% on SWE-Bench Verified, and 54.8% on Tau-Bench Retail. On knowledge and reasoning evaluations, OpenAI reports 85.3% on MMLU, 71.5% on GPQA Diamond without tools, and 17.3% on Humanity's Last Exam with tools. Results vary substantially with reasoning effort and tool access, so these numbers should not be treated as universal production accuracy rates.

OpenAI's own comparison places gpt-oss-20b close to o3-mini on several evaluation categories, while the larger gpt-oss-120b generally has an advantage on broad knowledge and agentic workloads. citeturn1search0turn1search1

gpt-oss-20b vs gpt-oss-120b vs o3-mini

The practical distinction is deployment control. gpt-oss-20b is the better choice when developers need open weights, local/private execution, customization, or relatively modest hardware requirements. gpt-oss-120b is preferable when higher overall reasoning and knowledge performance justify a larger deployment footprint.

Limitations of gpt-oss-20b

gpt-oss-20b is text-only and does not natively accept image, audio, or video inputs. Its smaller parameter count also creates a performance gap on some knowledge-heavy and agentic evaluations compared with gpt-oss-120b. In addition, open-weight deployment transfers more operational responsibility to the developer: hosting, scaling, monitoring, safety controls, and runtime configuration are not handled in the same way as a fully managed proprietary API. citeturn0search0turn0search2

OpenAI also notes that open-weight models have a different safety profile from hosted models because developers can modify or fine-tune the weights. Production deployments should therefore add appropriate application-level safeguards and access controls.

Use cases for gpt-oss-20b

gpt-oss-20b is a strong fit for:

1. Local coding assistants where developers want reasoning and code generation without sending code to a proprietary hosted model.
2. Mathematical and technical reasoning where high reasoning effort can be enabled for difficult problems.
3. Private enterprise workloads that require deployment inside a controlled environment.
4. Tool-using agents that combine the model with function calling, Python execution, web search, or application-specific tools.
5. Fine-tuned domain assistants where open weights are more valuable than access to a managed proprietary model.
6. Resource-constrained inference where the approximately 16 GB memory target enabled by MXFP4 quantization is important.

Accessing gpt-oss-20b through CometAPI

CometAPI currently lists gpt-oss-20b-free as an OpenAI model and describes it as a 21B-parameter open-source MoE model with 3.6B active parameters and a 128K-class context. It is free to use and exposes the model through CometAPI's unified API. The CometAPI changelog states that the model follows the OpenAI chat standard format.

For CometAPI integration, the general workflow is to create a CometAPI key, use the CometAPI base URL, and send the model name in the request. CometAPI documents OpenAI-SDK-compatible integration and currently lists https://api.cometapi.com/v1 as the quickstart base URL.