OpenAI and NVIDIA Drive AI Innovation Forward With Newly Released Open Models Fine-Tuned for the Largest AI Inference Infrastructure in the World
In a historic step for the worldwide AI community, OpenAI and NVIDIA have introduced two new open-weight AI reasoning models — gpt-oss-120b and gpt-oss-20b — granting unmatched access to state-of-the-art AI capability to developers, organizations, startups, governments, and inventors across the globe.
At the center of this work is NVIDIA’s industry-leading Blackwell GB200 NVL72 rack-scale system, which processes 1.5 million tokens per second on a single gpt-oss-120b instance — a new AI inference performance benchmark.
Open Models of the Coming AI Revolution
OpenAI’s latest open-weight text-reasoning large language models (LLMs) were trained on NVIDIA H100 GPUs and achieve best-in-class inference performance on the NVIDIA CUDA platform, which powers hundreds of millions of GPUs globally.
Now available immediately as NVIDIA NIM microservices, the models may be easily deployed across GPU-accelerated infrastructure, delivering flexibility, data privacy, and enterprise-grade security.
Optimized for the NVIDIA Blackwell platform, the models redefine inference efficiency, enabling breakthroughs across sectors including generative AI, physical AI, healthcare, manufacturing, and emerging industrial applications.
“OpenAI demonstrated the world what could be built on NVIDIA AI — and now they're pushing innovation forward in open-source software,” said Jensen Huang, NVIDIA founder and CEO. “The gpt-oss models enable developers around the globe to build on top of that leading-edge open-source foundation, enhancing U.S. tech leadership in AI — across the globe's largest AI compute infrastructure.”
NVIDIA Blackwell: Designed Specifically for Advanced Reasoning
As next-gen reasoning models generate exponentially more tokens, compute demand goes through the roof. NVIDIA Blackwell was built to respond to this challenge and boasts NVFP4 4-bit precision to deliver ultra-efficient, high-precision inference and reduced power and memory consumption.
This spike in efficiency allows scaling trillion-parameter LLM into production and unleashes vast commercial and industrial potential.
Empowering the Worldwide Developer Community
Now with over 450 million NVIDIA CUDA downloads, gpt-oss releases advanced AI tools into the hands of one of the world’s largest development ecosystems.
OpenAI and NVIDIA have cooperated with top open-source framework providers to provide widespread compatibility and best-in-class performance across FlashInfer, Hugging Face, llama.cpp, Ollama, vLLM, and NVIDIA TensorRT-LLM, and more — freeing developers to develop with their tools of choice.
History of Mutual Working under Open Innovation
The OpenAI–NVIDIA partnership traces back to 2016, when Huang personally brought the first NVIDIA DGX-1 AI supercomputer to OpenAI’s offices in San Francisco. From then on, the two organizations have cooperated on massive-scale training runs, ever pushing the boundaries of AI research and AI applications.
Today’s launch marks another milestone in this journey — an intersection of NVIDIA’s full-stack technology and OpenAI’s leading AI research — and the acceleration of worldwide innovation. Optimized for Blackwell and RTX GPUs, these open models will drive much-faster, more economical AI development among 6.5 million developers across 250 countries, using NVIDIA’s 900+ SDKs and AI models.
In short: This concurrent launch marks a watershed moment in AI access and capability, where OpenAI’s open-weight reasoning models meet NVIDIA’s unmatched inference infrastructure — driving the next era of innovation across industries, geographies, and communities.