Thread by @OpenAI

We released two open-weight reasoning models—gpt-oss-120b and gpt-oss-20b—under an Apache 2.0 license.

Developed with open-source community feedback, these models deliver meaningful advancements in both reasoning capabilities & safety.

openai.com/index/introduc…

gpt-oss-120b matches OpenAI o4-mini on core benchmarks and exceeds it in narrow domains like competitive math or health-related questions, all while fitting on a single 80GB GPU (or high-end laptop).

gpt-oss-20b fits on devices as small as 16GB, while matching or exceeding OpenAI o3-mini.

These models are trained for agentic workflows—supporting function calling, web search, Python execution, configurable reasoning effort, and full raw chain-of-thought access. github.com/openai/gpt-oss

We trained the gpt-oss models for reasoning, efficiency, and real-world usability across a wide range of deployment environments.

Both models were post-trained using our harmony response format to align them with the OpenAI Model Spec and teach them to apply chain‑of‑thought reasoning and tool use before producing answers.

github.com/openai/harmony

The gpt-oss models leverage our state-of-the-art approaches to safety training. We used deliberative alignment and an instruction hierarchy during post-training to help the models refuse unsafe prompts and defend against prompt injections, as well as pre-training interventions.

openai.com/index/gpt-oss-…

We adversarially fine-tuned gpt-oss-120b and evaluated the model. We found that even with robust fine-tuning, the model was unable to achieve High capability under our Preparedness Framework.

Our methodology was reviewed by external experts, marking a step toward new safety standards for open-weight models.

openai.com/index/estimati…

Both gpt-oss models are free to download on Hugging Face, with native MXFP4 quantization built in for efficient deployment.

Full list of day-one support is available on our blog.

huggingface.co/collections/op…