OpenAI publishes open-weight models for the first time since GPT-2

OpenAI released gpt-oss-120b and gpt-oss-20b, two mixture-of-experts language models distributed under the permissive Apache 2.0 licence — the first time the company had published model weights since GPT-2 in 2019, six years and several major releases earlier. The larger model has 117 billion total parameters with 5.1 billion active per token and was designed to run on a single 80GB GPU; the smaller has 21 billion total parameters with 3.6 billion active and was built to fit within 16GB of memory, putting it within reach of consumer hardware. Both support chain-of-thought reasoning, tool use, structured outputs and a configurable “reasoning effort” setting.

OpenAI reported that gpt-oss-120b matched or surpassed its own proprietary o4-mini model on core reasoning benchmarks including MMLU, and performed strongly on coding, health and agentic evaluations. Before release, the company said it had adversarially fine-tuned the models on biology and cybersecurity data specifically to test whether they could be pushed to dangerous capability levels, and reported that they did not reach those thresholds even under that adversarial training — an unusual step for an open-weight release, where a lab cannot retract or patch a model once its weights are public. OpenAI also launched a $500,000 red-teaming prize to surface further safety issues after launch.

The release reversed a position OpenAI had held since GPT-2’s staged release and GPT-3’s fully closed one: that frontier or near-frontier capability was too risky, or too commercially valuable, to publish as weights. By 2025 open-weight competition from Meta’s Llama family and Chinese labs including DeepSeek and Alibaba had made closed release look like a strategic liability rather than purely a safety choice, and gpt-oss gave OpenAI a stake in the open-weight ecosystem — and a benchmark for other labs to beat — without changing its approach to its flagship, closed frontier models.