OpenAI has released two new Artificial Intelligence (AI) models, gpt-oss-120b and gpt-oss-20b, which are freely downloadable and customizable by developers.
The new open-weight language models are now available on Hugging Face, a platform widely used by AI developers. OpenAI described the models as ‘state of the art,’ citing strong performance across industry benchmarks for open models.
The move to launch customizable AI models has been seen as a strategic response to growing pressure from competitors and the broader tech community for more transparency and openness.
Both AI models are released under the Apache 2.0 license, one of the most permissive open-source licenses, allowing commercial use without royalties or additional permissions. The models will be available to run on PCs through programs such as LM Studio and Ollama.
Cloud providers Amazon, Baseten and Microsoft are also making the models available. However, OpenAI confirmed it will not release the training data, citing ongoing copyright litigation within the AI industry.
The models come in two configurations: the more advanced gpt-oss-120b, designed to run on a single Nvidia GPU, and the lighter gpt-oss-20b, suitable for laptops with 16GB of memory.
This marks OpenAI’s first open language model release since GPT-2, more than five years ago. Since then, the company has favored a closed, proprietary approach to model development, monetizing its advanced AI systems through API services for businesses and developers.
OpenAI said that both customizable AI models were trained using similar processes to its proprietary models. Each model uses a mixture-of-experts (MoE) architecture, allowing only a subset of parameters to be activated per token. In the case of gpt-oss-120b, which features 117 billion total parameters, only 5.1 billion are used at a time, optimizing efficiency.
On Codeforces, a competitive programming benchmark, gpt-oss-120b and gpt-oss-20b scored 2622 and 2516, respectively. These results surpass DeepSeek’s R1 model but fall short of OpenAI’s own o3 and o4-mini models.
These customizable AI models were also trained with high-compute reinforcement learning (RL), similar to the method used in training OpenAI’s proprietary o-series models. This approach helps the AI develop a chain-of-thought reasoning process and makes the models well-suited for use in AI agents.
While capable of executing tools such as web search or Python code during responses, the new models are limited to text-only input and output, lacking capabilities in image or audio processing.
OpenAI delayed the release of customizable AI models to address potential safety risks. Internal and third-party testing indicated that the models could marginally improve biological capabilities, but did not meet the company’s ‘high capability’ threshold for misuse in areas such as cyberattacks or weapons development.
