August 6, 2025: OpenAI has made two new open-weight AI models, GPT OSS-120b and GPT OSS-20b, available on platforms like Hugging Face, Databricks, Azure, and AWS. The Apache 2.0 license lets anyone use the models for business purposes.
The GPT OSS-120b model can run on one Nvidia GPU and is as fast as OpenAI’s o4-mini model. The smaller 20b version is more efficient, needing only 16GB of RAM and working like the o3-mini model. Both models use a Mixture-of-Experts (MoE) architecture, which only activates a small number of parameters during inference to make the computation more efficient and lower the latency.
When you use reinforcement learning, the models get even better at thinking and following directions after they’ve been trained. They can be used with OpenAI’s Responses API and allow chain-of-thought reasoning, which makes them suitable for a wide range of applications, from automating business operations to aiding customers.