GPT OSS is a family of open-weight language models released by OpenAI in August 2025. It marks a major shift for the company, which hadn’t released publicly accessible model weights since GPT‑2 back in 2019. This new series includes two models: gpt-oss-120b, a high-performance model optimized for single H100 GPUs with 80 GB of memory, and gpt-oss-20b, a smaller model that can run efficiently on devices with 16 GB of memory - including consumer laptops and edge devices. Both models are designed to be highly capable in tasks like reasoning, coding, and chain-of-thought problem-solving.
If you’re inspired by the rapid advancements in open-weight AI models and want to build the skills to contribute to this space, consider taking the Python for Applied Data Science and AI* course on Coursera. It’s a beginner-friendly way to strengthen your Python foundations with real-world applications in data science and AI.
The models are released under the permissive Apache 2.0 license, allowing for commercial use, modification, and redistribution with minimal restrictions. This opens the door for startups, researchers, and developers to build powerful applications without relying on closed APIs or cloud infrastructure. OpenAI also emphasized safety, testing the models under its Preparedness Framework and finding them safe even when adversarially fine-tuned, especially in sensitive domains like cybersecurity and biothreats.
Technically, GPT OSS models use Mixture-of-Experts (MoE) architecture paired with MXFP4 quantization, which allows them to perform well while using far fewer active parameters per token (only 3–5 billion, depending on the model). Despite their compactness during inference, they show strong performance across standard benchmarks. Features like adjustable reasoning levels, tool use capabilities, and support for long-context processing (up to 128k tokens) make them suitable for building intelligent agents and complex workflows.
OpenAI has made GPT OSS available on major platforms such as Azure AI Foundry, AWS Bedrock, and Databricks, making it easier to deploy in enterprise environments. This release signals a big move toward democratizing access to advanced AI models, offering a level of transparency and control that aligns more closely with the open-source ecosystem while maintaining strong performance and safety standards.