OpenAI releases gpt-oss-120b and gpt-oss-20b — open-weights reasoning models under Apache 2.0, downloadable today. It is the company’s first open-weights language model since GPT-2 in 2019, the release whose staged rollout practically invented the “too dangerous to release” genre.

The models are mixture-of-experts reasoners with adjustable thinking effort, full chain-of-thought access, and tool use. OpenAI’s claim: gpt-oss-120b approaches o4-mini performance while fitting on a single 80GB GPU; the 20B variant runs on consumer hardware. Both went through the company’s Preparedness Framework, including deliberate adversarial fine-tuning to test worst-case misuse.

After months of delay — Sam Altman had publicly admitted being “on the wrong side of history” about open models — the timing reads as strategic on every axis: a year of Chinese open-weights releases setting the pace, a developer ecosystem drifting toward Qwen and Kimi, and, per the rumor mill, a rather large closed-model announcement expected later this week.