gpt-oss is a pair of open-weight reasoning models from OpenAI, released in August 2025 under the Apache 2.0 licence.
In plain words
gpt-oss comes in two sizes, gpt-oss-120b and gpt-oss-20b. Their weights are free to download, so anyone can run them on their own hardware. Both use a mixture-of-experts design, where only part of the network works on each token, and ship with their weights compressed in a 4-bit format called MXFP4. That lets the larger one fit on one 80 GB graphics card and the smaller one run within 16 GB of memory. You can set how hard they reason.
A quick example
Imagine a developer who cannot send data to the cloud. They download gpt-oss-20b from Hugging Face, run it on a laptop with enough memory, and set its reasoning effort to low for quick answers.
Where you'll meet it
- Local and on-device use.
- Private or offline deployments.
- Fine-tuning on specialised datasets.
Related
Sources
- DocsIntroducing gpt-oss, OpenAI · read 29 Sept 2026
- Papergpt-oss-120b & gpt-oss-20b Model Card, OpenAI (arXiv) · read 29 Sept 2026
- Docsgpt-oss, OpenAI (GitHub) · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.