OpenAI and NVIDIA release two new open-weight AI reasoning models, gpt-oss-120b and gpt-oss-20b, accessible to developers worldwide in various industries. The models were trained on NVIDIA H100 GPUs and run best on NVIDIA CUDA platform. NVIDIA Blackwell platform optimizations enable 1.5 million tokens per second inference efficiency.

NVIDIA Blackwell architecture supports advanced reasoning models like gpt-oss, utilizing innovations like NVFP4 4-bit precision for efficient and accurate low-precision inference. This architecture allows trillion-parameter LLMs to be deployed in real-time, unlocking significant value for organizations globally.

NVIDIA CUDA, the world’s most available computing infrastructure, provides access to the latest models for over 450 million developers. OpenAI and NVIDIA collaborate with top open framework providers to optimize models for various libraries, ensuring developers can build with their preferred framework.

NVIDIA’s collaboration with OpenAI dates back to 2016, highlighting a history of pushing AI boundaries together. Optimizing gpt-oss models for NVIDIA Blackwell and RTX GPUs, along with the extensive software stack, accelerates AI advancements for 6.5 million developers worldwide, leveraging over 900 NVIDIA software development kits and AI models.

Read more at NVIDIA: OpenAI and NVIDIA Propel AI Innovation With New Open Models Optimized for the World’s Largest AI Inference Infrastructure