SAN FRANCISCO – August 6, 2025 – OpenAI today announced the release of two open-source reasoning models utilizing a Mixture of Experts (MoE) architecture: GPT-OSS-120B and GPT-OSS-20B.
These two models are released under the Apache 2.0 open-source license, allowing developers and researchers to use, modify, and distribute them for free.
The new models focus on Chain-of-Thought reasoning, aiming to improve the accuracy of results by breaking complex tasks down into smaller steps. According to OpenAI's benchmarks, GPT-OSS-20B's performance is comparable to o4-mini, while GPT-OSS-120B's performance is comparable to o3-mini.
OpenAI CEO Sam Altman stated:
"We are thrilled to get these models into the hands of more people around the world. They will not replace existing GPT models, but will serve as a complement to drive innovation and collaboration in the artificial intelligence field."
A major highlight of this release is their ability to run on local devices. Smaller models like GPT-OSS-20B have been optimized to run efficiently on laptops and even smartphones. On an M4-powered MacBook, the 20B model can achieve inference speeds of 30-40 tokens per second.
Related Links:
- Model Info: https://openai.com/zh-Hans-CN/open-models/
- Model Download: https://huggingface.co/openai
Running GPT-OSS-20B Locally
On ollama, the GPT-OSS-20B model size is 14GB, and the one-click command to run it is as follows:
1 | ollama run gpt-oss:20b |
For a graphical user interface, you can use LM Studio
Comments