Microsoft brings GPU-accelerated gpt-oss-20B to Windows

OpenAI just dropped its open reasoning models, and Microsoft immediately brought a GPU-accelerated version of gpt-oss-20B to Windows devices for local offline inference.

It is built specifically for agentic workflows, meaning you can run local code execution and tool calling right on your own machine without sending data to cloud APIs.

What it can do

gpt-oss-20B is tuned for reasoning and tool use rather than general chatter. OpenAI trained it with high-compute reinforcement learning, which lets it break down tasks and call external tools, like web searches or Python scripts, directly inside its chain of thought.

Microsoft positions the model as lightweight and tool-savvy, built for power users who want autonomous assistants running in bandwidth-constrained setups.

Keep in mind that this is strictly a text-only model. It will not generate or process images or audio, keeping all of its horsepower focused on structured text tasks.

Specs, hardware, and halluncinations

To run it locally with decent speed, you will need a consumer desktop or laptop with at least 16GB of VRAM. A modern GPU from Nvidia or AMD Radeon fits the bill.

Before you jump in, you should also know its biggest weakness: facts about people.

On OpenAI's in-house PersonQA benchmark, which measures biographical factual accuracy, gpt-oss-20B hallucinated on 53% of the questions. That is a massive error rate if you plan to use it for knowledge retrieval or research.

Basically, do not treat it like Wikipedia. Treat it like a lightweight logic engine that drives scripts, and make sure you ground its outputs.

How to get it running

Microsoft made setup surprisingly painless through two routes.

The fastest CLI path is through Foundry Local. Open your Windows terminal and install it via WinGet by running winget install Microsoft.FoundryLocal. Once installed, just type foundry model run gpt-oss-20B to pull down the weights and start prompting from the command line.

If you prefer a graphical workflow, you can use the AI Toolkit extension inside Visual Studio Code. Search for AI Toolkit in extensions, open the Model Catalog, hit download on gpt-oss-20B, and open Model Playground to start tinkering with prompts and parameters.

Once you are done testing, both tools let you plug the local endpoint straight into custom apps using the Foundry Local SDK. Microsoft noted that macOS support is coming soon as well.

Why it matters

We think this release solves a very practical headache for developers who want private, local AI agents without paying per-token API fees. Most sub-30B local models struggle hard when juggling function calls, but a dedicated RL-trained reasoning model changes the math on consumer hardware.

If you have a 16GB VRAM GPU, it is definitely worth firing up for local scripting and automation. Just make sure you guard against its factual hallucinations and lean into what it is actually good at: running code and executing multi-step tools.