Windows 11 users now have access to OpenAI’s new free and open GPT model, gpt-oss-20b, through Windows AI Foundry, Microsoft’s platform for running AI tools, APIs, and popular open-source models directly on a PC.
Image Credit: The Economic Times
A Lightweight Model Built for Real Work
Microsoft describes gpt-oss-20b, in a blog post, as “tool-savvy and lightweight”. Translation? It’s designed to handle practical tasks like:
- Writing and debugging code
- Using tools to fetch information
- Running smoothly on different Windows devices
It doesn’t need super-expensive hardware either.
If you’ve got a GPU with at least 16GB of VRAM: modern Nvidia or AMD Radeon cards, you’re good to go.
The model uses high-compute reinforcement learning, which boosts its ability to power AI assistants and trigger actions like web searches or Python scripts.
This makes it ideal for developers building autonomous assistants or AI-powered workflows, even in places with slower internet.
What It Can and Can’t Do
Gpt-oss-20b is a text-only model.
That means:
- It can handle conversations, answer questions, and write code.
- It won’t process images, videos, or audio.
But there’s a catch: accuracy.
On OpenAI’s internal PersonQA benchmark, the model gave factually wrong answers 53% of the time when asked about people.
This “hallucination” rate means users should double-check sensitive information.
More Platforms Coming Soon
For now, gpt-oss-20b works on Windows PCs and laptops, but Microsoft says support for macOS and other devices is coming.
If you’d rather not run the model locally, it’s also available on:
- Microsoft Azure AI Foundry (hosted version)
- Amazon AWS
Comparing gpt-oss-20b and gpt-oss-120b
Microsoft is releasing two models side by side:
| Model | Size | Best For | Runs Locally? | Cloud Availability |
| gpt-oss-20b | Smaller | Local AI assistants, tool usage | Yes | Azure, AWS |
| gpt-oss-120b | Larger | More complex reasoning tasks | No | Azure, AWS |
For many Windows users, the smaller 20B model is easier to run without sacrificing too much capability for everyday AI tasks.
