Global collaboration between OpenAI and NVIDIA introduces expansive Open-Weight Reasoning Models, marking the advent of a scalable artificial intelligence era.
NVIDIA and OpenAI Collaborate to Advance AI Globally
NVIDIA and OpenAI have joined forces to revolutionize the field of artificial intelligence (AI) with the release of two new open-source models, gpt-oss-120b and gpt-oss-20b. These models, designed to deliver advanced reasoning capabilities, are optimized for NVIDIA's Blackwell architecture and are set to drive significant advancements in AI development worldwide.
Key Features
The new models boast several key features, including open-weight, flexible reasoning models, high performance and efficiency on NVIDIA hardware, large context handling, open-source accessibility, compatibility with popular AI frameworks, and training on industry-leading hardware.
Open-weight, Flexible Reasoning Models
Employing the mixture-of-experts (MoE) architecture with chain-of-thought reasoning and tool-calling capabilities, these models allow adjustable reasoning effort levels tailored to various tasks.
High Performance and Efficiency
Optimized for NVIDIA RTX AI GPUs and data center GPUs based on Blackwell and H100 architectures, the models deliver fast inference performance. For instance, gpt-oss-120b achieves up to 1.5 million tokens per second on NVIDIA GB200 NVL72 and 256 tokens per second on GeForce RTX 5090 GPUs.
Large Context Handling
Both models support a long input context length of up to 128k tokens using rotary positional embeddings (RoPE) that alternate between full context and sliding window attention.
Open-source and Broadly Accessible
The models are released as open-weight, enabling developers and enterprises worldwide to fine-tune, customize, and deploy AI applications without proprietary restrictions.
Compatibility with Popular AI Frameworks
NVIDIA and OpenAI have worked with frameworks such as Ollama, llama.cpp, Hugging Face Transformers, Microsoft AI Foundry Local, and NVIDIA TensorRT-LLM for optimized software stack integration.
Trained on Industry-leading Hardware
Both models underwent extensive training on NVIDIA H100 Tensor Core GPUs, with gpt-oss-120b requiring over 2.1 million GPU hours.
Benefits
The benefits of this collaboration are far-reaching, including democratizing AI development, advanced reasoning and domain capabilities, scalability and deployment flexibility, accelerated AI innovation and sovereignty, and cross-platform optimization.
Democratizing AI Development
By releasing powerful models openly and optimizing them across hardware, this collaboration lowers barriers for AI developers and researchers, fostering innovation beyond major organizations.
Advanced Reasoning and Domain Capabilities
Benchmark evaluations show gpt-oss-120b outperforms or matches many proprietary OpenAI models in coding, problem-solving, health queries, and math competitions, providing state-of-the-art performance for a broad range of tasks.
Scalability and Deployment Flexibility
Thanks to the open architecture and hardware optimization, these models run efficiently on edge devices, local AI PCs, and large cloud deployments, meeting diverse performance and regulatory requirements globally.
Accelerated AI Innovation and Sovereignty
The collaboration strengthens AI compute infrastructure and supports a broad ecosystem of software tools, enabling enterprises across industries (finance, healthcare, etc.) to adopt AI models compliant with local regulations.
Cross-platform Optimization
Partnership with AMD and Groq alongside NVIDIA ensures the models are performant on diverse GPU architectures, enhancing accessibility to a wide hardware base.
How these models enable AI advancements worldwide
The NVIDIA-OpenAI collaboration bridges cutting-edge model research with robust, optimized hardware and software ecosystems, delivering open models at scale that empower developers globally. This encourages experimentation, customization, and rapid deployment of advanced AI applications in various fields such as coding assistance, scientific research, healthcare diagnostics, and agentic AI for intelligent web search and tool use. The accessibility and efficiency achieved facilitate AI advancements beyond large tech companies, promoting innovation ecosystems worldwide.
In summary, the collaboration provides high-performance, versatile open models optimized for NVIDIA GPUs, which accelerate AI application development, democratize access, and enable scalable deployments globally across industries and devices. The release of these open models brings cutting-edge AI development directly into the hands of a wide range of users.
- The collaboration between NVIDIA and OpenAI delivers open-source AI models, such as gpt-oss-120b and gpt-oss-20b, that are optimized for NVIDIA's Blackwell architecture, democratizing AI development and fostering innovation beyond major organizations.
- The released models, featuring advanced features like large context handling and flexible reasoning models, are compatible with popular AI frameworks like Hugging Face Transformers and NVIDIA TensorRT-LLM, enabling developers to customize and deploy AI applications seamlessly.
- By training the models on industry-leading hardware like NVIDIA H100 Tensor Core GPUs, the collaboration ensures high performance and efficiency, making advanced AI applications viable on edge devices, local AI PCs, and large cloud deployments, accommodating diverse performance and regulatory requirements worldwide.
