OpenAI and NVIDIA have expanded their collaboration with the release of two open-weight language models, gpt-oss-120b and gpt-oss-20b, aimed at democratizing advanced Artificial Intelligence development across industries and user groups. These models make high-level reasoning and generative tasks accessible to developers, startups, enterprises, and governments globally, marking a significant step toward community-driven technological innovation.
The models were developed using NVIDIA H100 GPUs and are specifically optimized for global deployment on the widely adopted NVIDIA CUDA platform, which supports hundreds of millions of GPUs worldwide. NVIDIA is making the models available as NVIDIA NIM microservices, simplifying deployment on any GPU-accelerated infrastructure, and addressing key requirements for flexibility, enterprise-grade security, and data privacy. Further optimizations for NVIDIA´s Blackwell architecture and GB200 NVL72 systems unlock industry-leading inference speeds of 1.5 million tokens per second, reducing operational costs and supporting real-time deployment of trillion-parameter models.
With over 450 million CUDA downloads, the integration of these models allows NVIDIA´s expansive developer community to immediately leverage them across various platforms, from DGX Cloud servers to personal GeForce RTX devices. OpenAI and NVIDIA have actively worked with open-source frameworks, ensuring compatibility with tools like FlashInfer, Hugging Face, llama.cpp, Ollama, vLLM, and NVIDIA´s TensorRT-LLM, enabling flexibility in developer workflows. The companies´ collaboration, which dates back to 2016, has continually pushed the boundaries of Artificial Intelligence scalability and accessibility. By bringing these new models to the broader ecosystem and optimizing them for both existing and next-generation NVIDIA hardware, OpenAI and NVIDIA underscore their commitment to open innovation and accelerating Artificial Intelligence-powered transformation worldwide.
