OpenAI Releases Open-Weight AI Models GPT-OSS-120B and GPT-OSS-20B Under Apache 2.0 License

OpenAI has unveiled two new open-weight language models, GPT-OSS-120B and GPT-OSS-20B, both text-only and released under an Apache 2.0 license. The models are available for download on platforms like GitHub and Hugging Face, with compatibility extending to LM Studio and Ollama. Designed for versatility, they can operate on hardware ranging from consumer laptops to cloud servers, supporting chips from Nvidia, AMD, Cerebras, and Groq.

Safety measures were a priority in the release process. OpenAI conducted rigorous evaluations to filter sensitive data—particularly chemical, biological, radiological, and nuclear content—during pre-training. The company also tested potential misuse scenarios, ensuring the models resisted malicious fine-tuning attempts. None of these tests triggered OpenAI's internal 'high capability' threshold for harmful potential, as defined by its Preparedness Framework.

'It’s been exciting to see an ecosystem develop,' said Greg Brockman, OpenAI’s president, emphasizing the company's commitment to advancing the field. Sam Altman highlighted the breakthrough, noting the larger model rivals closed alternatives like GPT-4 Mini in performance while running on high-end laptops, with a smaller variant even operable on smartphones.