Microsoft has made OpenAI’s new small open-weight language model available to Windows 11 users. The model, called gpt-oss-20b, is now accessible through Microsoft’s Windows AI Foundry tool. This release enables users to run a capable AI model directly on their own computers, eliminating the need for cloud services.
What is gpt-oss-20b?
OpenAI released gpt-oss-20b on August 5, 2025, alongside a larger version called gpt-oss-120b. Both models are open-weight, which means anyone can download and run them without connecting to OpenAI’s servers. The smaller 20-billion-parameter version is designed to run on modern consumer hardware. It can perform text-based tasks such as answering questions, writing drafts, or following instructions. It cannot process images or audio.
The gpt-oss-20b model focuses on what OpenAI calls “agentic” use cases. These include connecting the model to tools such as code execution environments or search engines. This allows the model to complete multi-step tasks with less human input.
Why it matters for Windows users
By bringing GPT-OS-20 b to Windows AI Foundry, Microsoft is giving developers and advanced users the ability to run AI workloads locally. This can lower latency and keep data on the user’s device, which is useful for privacy-sensitive projects. It also reduces dependence on internet connections and cloud subscriptions.
Running AI locally has limits. The GPT-OS-20 b model still requires a modern graphics card with at least 16 GB of video memory. Users without this hardware will not be able to run the model effectively. The larger gpt-oss-120b model is available only through Azure AI Foundry for cloud-based use.
Performance and limitations
Although GPT-OS-20 b is accessible and runs locally, it does not match the accuracy of OpenAI’s most advanced proprietary models. In internal tests, it showed a relatively high rate of incorrect answers in some benchmarks. For example, OpenAI’s PersonQA test found that the model produced wrong responses in more than half of the cases. Users will need to review and verify its outputs, especially when accuracy is essential.
The model is also text-only. It cannot process images, audio files, or video. It is best suited for structured tasks where inputs and outputs remain in plain text.
Microsoft’s broader approach
This release fits into Microsoft’s broader “hybrid AI” strategy. Users can start with a model running locally and, when needed, move heavier workloads to the cloud through Azure AI Foundry. This flexibility helps developers choose the right balance between performance, cost, and privacy.
Microsoft plans to expand support for gpt-oss-20b to more devices in the future, including Copilot+ PCs and possibly other operating systems such as macOS.
A step toward local AI
The arrival of GPT-OS-20 B in Windows AI Foundry marks a significant shift in how AI can be utilized. For the first time in years, OpenAI has provided an open-weight model that runs directly on a personal computer without a cloud connection. While it is not the most advanced AI available, its availability and openness will likely encourage more experimentation, learning, and custom development on Windows.
and then