GPT-OSS 20B

GPT-OSS 20B is one half of OpenAI's first openly released reasoning family, introduced on August 5, 2025 alongside the larger gpt-oss-120b. Both models were developed with input from the open-source community and are distributed under the Apache 2.0 license together with OpenAI's gpt-oss usage policy, making the full weights freely available for local deployment, fine-tuning, and downstream research. The accompanying technical report is published on arXiv as 2508.10925, giving practitioners a detailed reference for the architecture and evaluation choices behind the release. Because the weights are downloadable, teams can integrate the model into self-hosted inference stacks and tailor it to private or regulated environments where API-only access is not an option.

Designed as a text-only model, GPT-OSS 20B targets agentic workloads that combine reasoning, instruction following, and tool use. OpenAI highlights strong instruction adherence, native support for web search and Python code execution, and the ability to tune the reasoning effort so lighter tasks do not pay the latency cost of deep deliberation. The model exposes its full chain-of-thought for inspection, supports Structured Outputs for predictable downstream parsing, and is compatible with the Responses API, which makes it straightforward to drop into existing OpenAI-style pipelines. In practice, it fits well for developers building assistants that need transparent reasoning traces, controllable depth of thinking, and the freedom to self-host an OpenAI-aligned model on their own infrastructure.