gpt-oss-20b
OpenAI · United States · 2025
The smaller half of OpenAI’s open-weight pair — 21B parameters, meant to run locally rather than in a data centre.
gpt-oss-20b was released on 5 August 2025 alongside its larger sibling gpt-oss-120b, and shares its architecture, licence and design decisions at roughly a fifth of the size: about 21 billion parameters, of which 3.6 billion are active per token. OpenAI positions it for low latency, local execution and specialised fine-tunes rather than for frontier reasoning. The specification otherwise mirrors the larger model. The context window is 131,072 tokens with the same figure available on output, the knowledge cutoff is 1 June 2024, reasoning effort is selectable across low, medium and high, and the complete chain of thought is returned. Function calling, browsing, Python execution and structured outputs all work natively, and the weights are fine-tunable. The licence is Apache 2.0. As with gpt-oss-120b, OpenAI does not serve this model on its own API — the published rate limits are zero on every tier — so it reaches users through third-party hosts or through local runtimes on consumer hardware, which is the case its size was chosen for.
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!