Technical Specifications and Performance

The gpt-oss-120b model delivers performance close to OpenAI's o4-mini on core reasoning tasks, all while running efficiently on a single 80 GB GPU, a notable achievement in model optimization. Meanwhile, the gpt-oss-20b matches the performance of OpenAI's o3-mini across common benchmarks and is optimized for environments with limited resources, running smoothly on edge devices with as little as 16 GB of memory. This makes it well-suited for on-device applications, local inference, and fast prototyping without the need for expensive infrastructure.

In addition, both models demonstrate strong capabilities in areas such as tool usage, few-shot function calling, and chain-of-thought (CoT) reasoning as evidenced by evaluations on Tau-Bench and HealthBench. Notably, their performance in these areas even exceeds some proprietary models, including OpenAI o1 and GPT-4o.

But this isn't just about access to models, it's about redefining how collaboration in AI works. With the introduction of Open Models, OpenAI is empowering the developer and research communities to:

• Freely explore and build upon cutting-edge AI systems
• Contribute improvements, create custom applications, and accelerate innovation
• Participate in transparent evaluation and development processes

By making select fine-tuned and optimized models openly available, OpenAI is setting a new industry standard for openness, accountability, and technical excellence. This move enables stakeholders, enterprises, and communities alike to work together in enhancing AI responsibly and at scale.

This release underscores a larger vision - AI that is more accessible, trustworthy, and shaped by the broader community.