OpenAI has released gpt-oss-120b and gpt-oss-20b, its first open-weight language models since 2019. The models are made available under the Apache 2.0 license, a permissive license that allows modification and commercial use. This marks a shift from OpenAI's previous focus on proprietary models.
The new models employ a mixture-of-experts (MoE) architecture. The gpt-oss-120b model, with 117 billion parameters, is designed for high-reasoning tasks and can run on a single 80GB GPU. The gpt-oss-20b model, with 21 billion parameters, is optimised for consumer-grade hardware and edge devices with at least 16GB of memory. Both models support tool use and multi-step reasoning.
These models are integrated into platforms like Azure AI Foundry and Windows AI Foundry, and are also available on Hugging Face. They can be fine-tuned and deployed without API dependencies, offering developers greater control. OpenAI has also implemented safety measures, including comprehensive safety training and evaluations, and is offering a reward for identifying novel safety issues.
