OpenAI on Tuesday announced the launch of two reasoning AI models in two sizes: a larger and more powerful gpt-oss-120b model that can run on a single Nvidia GPU, and a lighter gpt-oss-20b model that can run on a consumer laptop with 16GB of memory.
OpenAI on Tuesday announced the launch of two reasoning AI models in two sizes: a larger and more powerful gpt-oss-120b model that can run on a single Nvidia GPU, and a lighter gpt-oss-20b model that can run on a consumer laptop with 16GB of memory.
This is the first launch of an open OpenAI model since GPT-2, which was released more than 5 years ago, TechCrunch writes . They are freely downloadable from the online developer platform Hugging Face.
OpenAI said in a briefing that its open models can send complex queries to AI models in the cloud. This means that if an open model can’t perform a certain task, such as image processing, developers can connect it to one of the company’s more powerful closed models.
In Codeforces, a coding test, gpt-oss-120b and gpt-oss-20b scored 2622 and 2516 points respectively, outperforming R1 DeepSeek but falling behind o3 and o4-mini.
On Humanity's Last Exam (HLE), a complex test consisting of crowdsourced questions on various topics, gpt-oss-120b and gpt-oss-20b scored 19% and 17.3%, respectively. These results are lower than o3 but outperform the leading open-source models from DeepSeek and Qwen.
It is noteworthy that OpenAI's open models cause significantly more hallucinations than the latest AI thinking models, o3 and o4-mini.
OpenAI found that gpt-oss-120b and gpt-oss-20b hallucinated in response to 49% and 53% of questions, respectively, on the PersonQA test, the company’s own benchmark for measuring the accuracy of a model’s knowledge of people. This is more than three times the hallucination rate of OpenAI’s o1 model, which scored 16%, and higher than the o4-mini model, which scored 36%.
OpenAI claims that its open models were trained using similar processes to its own models. Each open model uses a Mixture-of-Experts (MoE) architecture to engage fewer parameters for a given question, making it more efficient. For example, in the gpt-oss-120b model, which has a total of 117 billion parameters, OpenAI says that for each token, the model only activates 5.1 billion parameters.
The company also says its open model was trained using high-performance reinforcement learning (RL) — a process that occurs after prior training to teach AI models to distinguish right from wrong in simulated environments — using large clusters of Nvidia GPUs. This method was also used to train OpenAI’s o-series models, and open models have a similar “chain-of-thought” process, where they spend additional time and computing resources to think through their answers.
Its open AI models are great for supporting AI agents and can invoke tools like web searches or Python code execution as part of their thought process. However, OpenAI says its open models are only available in text format, meaning they won’t be able to process or generate images and audio like the company’s other models.
OpenAI is releasing gpt-oss-120b and gpt-oss-20b under the Apache 2.0 license, which is generally considered one of the most permissive licenses. This license will allow businesses to monetize OpenAI’s open models without paying for them or getting permission from the company.
However, unlike fully open offerings from artificial intelligence labs such as AI2, OpenAI says it will not publish the training data used to build its open models.
