OpenAI Releases Its First Open-Source MoE Model

In a significant strategy shift, OpenAI has released gpt-oss-120b, its first-ever open-weight foundation model. The move places the historically closed-source company in direct competition with established open-source players like Meta and Mistral AI, signaling a new phase in the development of freely available AI.

The new model is a Mixture-of-Experts (MoE) architecture with a total of approximately 117 billion parameters. MoE models are designed for efficiency, activating only a fraction of their total parameters for any given task, which can lead to faster inference times and lower computational costs. According to the release notes, the model uses the MXFP4 data format and supports a generous 131,072-token context window.

OpenAI is releasing the model under the business-friendly Apache 2.0 license. This permissive approach allows for broad use, including commercial applications, which will likely accelerate its adoption among developers and researchers. The company highlights the model's strong performance on agentic and complex reasoning tasks, positioning it as a powerful tool for building sophisticated applications.

This release could reshape the open-source AI ecosystem by providing a new, high-quality baseline for development. Researchers and developers can now directly inspect, build upon, and fine-tune a model from the creators of the GPT series. All model weights and supporting files are available for download from the official Hugging Face repository.