Global collaboration between OpenAI and NVIDIA unveils Open-Weight Reasoning Models on a global scale, marking a substantial shift towards a new era of scalable artificial intelligence.
OpenAI, the leading artificial intelligence research lab, has recently released two new open-weight AI reasoning models: the gpt-oss-120b and gpt-oss-20b. In collaboration with NVIDIA, these models are set to push the boundaries of AI technology, providing the core technologies and expertise needed for massive-scale training runs.
The gpt-oss models are built on NVIDIA's AI compute infrastructure and offer optimal inference on NVIDIA's GB200 NVL72 systems, achieving 1.5 million tokens per second. The gpt-oss-120b, with 120 billion parameters, delivers strong reasoning, coding, and tool-use capabilities, while the gpt-oss-20b, with approximately 20 billion parameters, is more lightweight and efficient for edge devices.
These models are designed for developers, enthusiasts, enterprises, startups, and governments worldwide. They can be used to develop applications in various fields such as generative AI, reasoning AI, physical AI, healthcare, manufacturing, and potentially new industries.
NVIDIA's full-stack approach helps bring the world's most ambitious AI projects to the broadest user base possible. The gpt-oss models are API-compatible with common OpenAI interfaces, allowing easy integration and swap-in for existing applications.
With NVIDIA's collaboration, these models are now more accessible worldwide. They can be deployed on diverse hardware setups, including consumer-level GPUs, ensuring efficient inference and tool use across platforms. This partnership supports efficient inference and tool use across Windows, soon on MacOS, and in cloud instances like Azure AI Foundry. AMD has also enabled support for these models on Ryzen AI processors and Radeon GPUs.
The gpt-oss-120b is optimized for high-capacity, reasoning-heavy, autonomous AI tasks, while the gpt-oss-20b caters to lightweight, edge-friendly AI applications. NVIDIA's collaboration with OpenAI unlocks new possibilities for AI deployment with control, performance, and portability worldwide.
As of today, the CUDA developer community gains access to the latest models, optimized to run on the NVIDIA technology stack they already use. The models were trained on NVIDIA H100 GPUs and run inference best on the NVIDIA CUDA platform, which is globally distributed.
NVIDIA's Blackwell architecture, designed for advanced reasoning models, is crucial for meeting the increasing demand as these models generate more tokens. Deploying trillion-parameter LLMs in real time with NVIDIA Blackwell can unlock billions of dollars in value for organizations.
OpenAI and NVIDIA have collaborated with top open framework providers to offer model optimizations for FlashInfer, Hugging Face, llama.cpp, Ollama, and vLLM, in addition to NVIDIA Tensor-RT LLM and other libraries. This combined hardware-software ecosystem promises new possibilities for AI deployment with control, performance, and portability worldwide.
This collaboration between OpenAI and NVIDIA, which began in 2016 with the hand-delivery of the NVIDIA DGX-1 AI supercomputer to OpenAI's headquarters, continues to advance innovation in open-source software. Jensen Huang, Founder and CEO of NVIDIA, stated that OpenAI's use of NVIDIA AI technology has been instrumental in this advancement.
- The gpt-oss models, built on NVIDIA's AI technology and infrastructure, are designed to cater to diverse sectors, including healthcare, helping startups, enterprises, and governments worldwide develop applications leveraging AI technology.
- With the collaboration between OpenAI and NVIDIA, the gpt-oss-120b and gpt-oss-20b models are optimized for various use cases, such as cloud instances, edge devices, and hardware setups like consumer-level GPUs, ensuring worldwide accessibility.
- The combined hardware-software ecosystem developed by OpenAI, NVIDIA, and top open framework providers like FlashInfer, Hugging Face, and Ollama, offers model optimizations for efficient AI deployment, aiming to unlock new possibilities and deliver control, performance, and portability globally.
