OpenAI has launched two new open AI models capable of logical reasoning — gpt-oss-120b and gpt-oss-20b. These are the first open-weight models released in over five years since GPT-2. Both are freely available on Hugging Face, appealing to developers and researchers eager to create their own solutions based on open technologies.
The models differ in power and hardware requirements:
- gpt-oss-120b — a more robust model that can operate on a single NVIDIA GPU;
- gpt-oss-20b — a lighter version that can run on a standard laptop with 16 GB of RAM.
OpenAI's goal is to provide an American open AI platform as an alternative to the growing influence of Chinese laboratories such as DeepSeek, Qwen (Alibaba), and Moonshot AI, which are actively developing powerful open models.
In testing on the competitive coding platform Codeforces, the 120b model scored 2622 points, while the 20b model achieved 2516, surpassing DeepSeek R1 but falling short of closed models o3 and o4-mini. In the challenging Humanity’s Last Exam (HLE), the 120b reached 19%, and the 20b achieved 17.3%, outperforming other open models but still lagging behind o3.
The new models were trained using methodologies similar to those employed in OpenAI's closed models. They utilize a mixture-of-experts (MoE) approach, activating only a portion of parameters for each token, enhancing efficiency. An additional RL-fine-tuning allowed the models to learn to create chains of logical reasoning and invoke tools like web search or execute Python code.
The models operate solely with text and do not generate images or audio. They are distributed under the Apache 2.0 license, permitting commercial use without agreement from OpenAI, although the training data remains proprietary due to copyright risks.
The launch of gpt-oss aims to strengthen OpenAI's position within the developer community and respond to political pressure from the U.S. seeking to enhance the role of open American models in global competition.
