OpenAI and NVIDIA Unveil Open-Weight AI Models for Local Deployment on RTX GPUs

OpenAI and NVIDIA Collaborate on Open-Weight AI Models

In a groundbreaking move, OpenAI and NVIDIA have joined forces to release two new open-weight AI reasoning models, gpt-oss-20b and gpt-oss-120b. This collaboration marks a significant step forward in democratizing access to advanced AI technologies, allowing developers, enthusiasts, and organizations to run sophisticated language models locally on their own hardware 1

1

2

2

.

Model Specifications and Performance

The gpt-oss-20b model, designed for broader accessibility, can run on GPUs with at least 16GB of VRAM. It offers performance comparable to OpenAI's o3-mini model on common benchmarks 3

3

. For more demanding applications, the gpt-oss-120b model achieves near-parity with OpenAI's o4-mini on core reasoning benchmarks and requires an 80GB GPU 3

3

.

NVIDIA has optimized these models for their hardware, showcasing impressive performance metrics:

• On a NVIDIA GeForce RTX 5090 GPU, the gpt-oss-20b model can process up to 256 tokens per second 1

  1

  .
• The gpt-oss-120b model, when run on NVIDIA Blackwell GB200 NVL72 systems, achieves an astounding 1.5 million tokens per second 2

  2

  .

Deployment Options and Software Stack

Users have multiple options for deploying these models locally:

1. Ollama: A user-friendly application that provides out-of-the-box support for OpenAI's open-weight models, optimized for RTX GPUs 1

   1

   .
2. Microsoft AI Foundry Local: Currently in public preview, this solution integrates into workflows via command line, SDK, or APIs 1

   1

   .
3. llama.cpp: An open-source framework optimized for RTX GPUs, featuring recent contributions like CUDA Graphs implementation 1

   1

   5

   5

   .

Implications for AI Development and Industry

The release of these open-weight models under the Apache 2.0 license allows for full commercial and research use, potentially accelerating AI innovation across various sectors 3

3

. Jensen Huang, founder and CEO of NVIDIA, emphasized the significance of this release:

"OpenAI showed the world what could be built on NVIDIA AI -- and now they're advancing innovation in open-source software. The gpt-oss models let developers everywhere build on that state-of-the-art open-source foundation, strengthening U.S. technology leadership in AI -- all on the world's largest AI compute infrastructure." 2

2

Accessibility and Hardware Requirements

While the gpt-oss-20b model is accessible to a wider range of users with RTX GPUs featuring at least 16GB of VRAM, the more powerful gpt-oss-120b model requires more substantial hardware. AMD has also announced support for these models, with CEO Lisa Su confirming compatibility with AMD AI CPUs and GPUs 4

4

.

Privacy and Local Deployment Benefits

Running these models locally offers several advantages, including enhanced privacy, reduced dependence on cloud services, and the ability to work offline. This makes the technology particularly attractive for sectors like finance, healthcare, and government, where data sensitivity is a primary concern 5

5

.

As AI continues to integrate into various aspects of computing and industry, the release of these open-weight models by OpenAI and NVIDIA represents a significant milestone in making advanced AI capabilities more accessible and customizable for developers and organizations worldwide.