Microsoft Brings OpenAI's Smallest Model to Windows

Microsoft Integrates OpenAI's gpt-oss-20b into Windows AI Foundry

Microsoft is now providing access to OpenAI’s recently released, free and open-source gpt-oss-20b model for Windows 11 users. This integration is facilitated through Windows AI Foundry, the company’s platform designed to enable users to utilize AI capabilities.

Accessing AI Features and Models

Windows AI Foundry allows users to leverage AI features, APIs, and a variety of popular open-source models directly on their computers. The availability of gpt-oss-20b expands the options available within this platform.

gpt-oss-20b: Characteristics and Capabilities

According to a company blog post, gpt-oss-20b is characterized as being both “tool-savvy and lightweight.” It is specifically optimized for tasks requiring agentic behavior, such as code execution and tool utilization.

The model is designed to operate efficiently across a diverse range of Windows hardware configurations. Further support for additional devices is anticipated in the near future.

This makes it well-suited for the development of autonomous assistants or the integration of AI into practical, real-world workflows, even in situations with limited bandwidth.

Technical Specifications and Training

OpenAI launched gpt-oss-20b on Tuesday, noting its ability to function on consumer PCs and laptops equipped with at least 16GB of VRAM. This requirement is commonly met by modern GPUs from manufacturers like Nvidia or Radeon.

The model’s training process involved the use of high-compute reinforcement learning. This technique enhances its performance in powering AI agents and facilitating tool calls, including functionalities like web searches and Python code execution, as part of its reasoning process.

Limitations of the Model

It’s important to note that gpt-oss-20b is currently a text-only model. Consequently, it lacks the ability to process or generate images and audio, unlike some of OpenAI’s other offerings.

Accuracy and Hallucination Rate

The model exhibits a tendency towards hallucination. Testing on OpenAI’s internal benchmark, PersonQA, revealed that gpt-oss-20b produced inaccurate responses to 53% of questions related to factual knowledge about individuals.

Future Availability and Platform Support

Microsoft plans to extend support for the model to macOS in the coming months, alongside other unspecified devices. Both gpt-oss-20b and the older gpt-oss-120b model are also accessible through Microsoft’s hosted Azure AI Foundry platform.

Availability on AWS

In addition to Microsoft’s platforms, these two new models are also available for use on Amazon’s AWS.