gpt-oss-120b

gpt-oss-120b, published on 5 August 2025, ended a six-year stretch in which OpenAI released no downloadable weights at all — the last had been GPT-2 in 2019. It is a mixture-of-experts transformer with roughly 117 billion parameters, of which about 5.1 billion are active per token, and OpenAI designed it to fit inside a single 80 GB H100 accelerator. The context window is 131,072 tokens and the model can emit the same number on output, an unusually symmetric arrangement; the knowledge cutoff is 1 June 2024. Reasoning effort is configurable across low, medium and high, and the full chain of thought is exposed rather than summarised — a deliberate choice OpenAI framed as an aid to debugging and oversight, and one its hosted models do not offer. The model calls functions, browses, runs Python and returns structured output, and it can be fine-tuned. The licence is Apache 2.0, with no copyleft obligation and no field-of-use restriction, so commercial deployment is permitted outright. OpenAI itself does not serve the model through its API — the published rate limits are zero across every usage tier — so anyone wanting it hosted rents it from a third-party provider or runs it themselves.

▸Videos

No videos yet.

▸Reviews