NVIDIA collaborates with OpenAI to optimize new open-source reasoning models for NVIDIA GPUs, enabling agentic AI applications like web search. The gpt-oss-20b and gpt-oss-120b models are now accessible to millions, offering up to 256 tokens per second performance on NVIDIA RTX AI PCs. Jensen Huang praises OpenAI for advancing innovation in open-source software.

The flexible gpt-oss models support up to 131,072 context lengths, ideal for web search, coding assistance, and document comprehension. Trained on NVIDIA H100 GPUs, these models are the first MXFP4 models supported on NVIDIA RTX, offering high model quality and efficient performance. Developers can test them using the Ollama app on RTX AI PCs.

Ollama’s new app features easy chatting with models, support for PDF or text files within chats, multimodal support, and customizable context lengths. Developers can also use Ollama via command line or SDK for their applications. Enthusiasts and developers can explore the gpt-oss models on RTX AI PCs through various applications and frameworks, all powered by RTX GPUs.

Windows developers can access OpenAI’s models via Microsoft AI Foundry Local, an on-device AI inferencing solution in public preview. Foundry Local uses ONNX Runtime, optimized through CUDA, with NVIDIA TensorRT for RTX support coming soon. The release of these open-source models marks a new wave of AI innovation for developers looking to enhance their AI-accelerated Windows applications.

Read more at NVIDIA: OpenAI’s New Models on RTX GPUs