OpenAI Unveils First Open-Weight AI Models in Five Years

OpenAI has unveiled its first open-weight AI models in over five years, marking a significant strategic pivot aimed at broadening developer access and influencing global AI policy. The release introduces two powerful reasoning models, gpt-oss-120b and gpt-oss-20b, now publicly available on Hugging Face under the highly permissive Apache 2.0 license. This move signifies OpenAI’s return to its open-source roots, a departure from its recent focus on proprietary, API-driven models.

The decision to release these open-weight models, meaning their underlying parameters are freely available for download, modification, and deployment, underscores a shift towards democratizing AI technology. This enables developers and organizations to customize, fine-tune, and run these models on their own infrastructure, reducing reliance on remote cloud APIs and enhancing data privacy. The Apache 2.0 license is particularly noteworthy, as it allows for unrestricted commercial use, modification, and distribution, including patent grants that protect users from infringement claims. This permissive licensing stands in contrast to some other open-source models, which may include restrictions on commercial use or scale.

The gpt-oss-120b model, with its 117 billion parameters, is designed for high-performance applications and achieves near-parity with OpenAI’s o4-mini on core reasoning benchmarks. Impressively, this larger model can run efficiently on a single 80GB Nvidia GPU. The lighter gpt-oss-20b, featuring 21 billion parameters, offers performance comparable to OpenAI’s o3-mini and is optimized for consumer hardware, capable of running on devices with just 16GB of memory, making it ideal for on-device use cases and local inference. This accessibility is a game-changer, breaking down barriers for smaller organizations, emerging markets, and individual developers who may lack the budget for extensive cloud infrastructure.

Both gpt-oss models excel in reasoning tasks, demonstrate strong tool-use capabilities, and support few-shot function calling and Chain-of-Thought (CoT) reasoning. They are compatible with OpenAI’s Responses API and designed for agentic workflows, offering adjustable reasoning efforts for varied tasks. While the models are highly capable, OpenAI has emphasized safety, implementing comprehensive training and evaluations to minimize risks, even under malicious fine-tuning. However, the Chain-of-Thought results are left unfiltered for transparency, which may lead to higher hallucination scores, a trade-off for monitorability.

This strategic shift by OpenAI comes amid increasing competition from other open-source AI initiatives, such as Meta’s LLaMA and Chinese models like DeepSeek, which have demonstrated the power and collaborative potential of open-weight approaches. OpenAI’s CEO Sam Altman previously acknowledged being “on the wrong side of history” regarding openness, and this release signals a renewed commitment to making AI broadly accessible. The company aims to foster innovation through open collaboration, believing that a wider developer base will accelerate research and lead to safer, more transparent AI development.

Furthermore, the models are now available on major cloud platforms like Amazon Bedrock and Amazon SageMaker JumpStart, expanding their reach to millions of AWS customers. Microsoft is also integrating GPU-optimized versions of gpt-oss-20b into Windows devices, further enhancing local deployment options. This widespread availability and the permissive licensing are set to empower a new wave of AI innovation, allowing builders to create and deploy AI solutions on their own terms, from the cloud to the edge.