TLDR: OpenAI has released its first open-source, open-weight AI models, gpt-oss-120b and gpt-oss-20b, since 2019. The larger model is now available on IBM watsonx.ai, marking a significant shift towards more accessible and customizable AI solutions for enterprises and developers.
In a landmark move signaling a return to its open-source roots, OpenAI has announced the release of two new open-weight artificial intelligence models: gpt-oss-120b and gpt-oss-20b. This marks the first time since 2019 that the AI titan has made its model weights freely available, allowing developers and businesses to download, run, and customize them without relying on remote cloud APIs or exposing sensitive in-house data to external services. This initiative aims to lower barriers for emerging markets, resource-constrained sectors, and smaller organizations that may find proprietary models less accessible.
One of these newly released models, gpt-oss-120b, is now hosted on IBM’s watsonx.ai developer studio, with the smaller gpt-oss-20b model slated for integration soon. This collaboration enables secure deployment within enterprise infrastructure, allowing businesses to customize the models using their internal data and embed them into real-world workflows across various industries, including healthcare, financial services, and legal support. IBM watsonx.ai, a unified development platform, empowers businesses to develop, deploy, and scale AI applications efficiently.
The gpt-oss models are designed for robust reasoning capabilities and tool usage, supporting features like web search and Python code execution within a chain-of-thought framework. They offer adjustable reasoning levels—low, medium, or high—allowing users to balance output quality against cost and speed. Benchmarking results highlight their strong performance: gpt-oss-120b scored 90 on the MMLU benchmark, 80.1 on the GPQA science benchmark, and an impressive 97.9 on the AIME 2025 math test, indicating strong symbolic reasoning. The models also output full chains of reasoning, enhancing transparency and making them easier to debug and trust.
Technically, the models are licensed under Apache 2.0, one of the most permissive open licenses, enabling broad use, modification, and commercialization. The gpt-oss-120b model is optimized to run efficiently on a single NVIDIA H100 GPU (requiring 80GB of memory), while the gpt-oss-20b is designed for consumer-grade devices like laptops with 16GB of memory. Both models employ a mixture-of-experts (MoE) architecture, contributing to their efficiency and performance.
OpenAI emphasizes safety as a foundational aspect of this release. Both models have undergone thorough safety training and evaluation, including rigorous testing of a maliciously fine-tuned version of gpt-oss-120b under OpenAI’s Preparedness Framework. These methods, reviewed by external safety experts, represent a meaningful advancement in open model safety standards.
Also Read:
- OpenAI Unveils New Open-Weight Models, Expanding Accessibility and Cloud Partnerships
- OpenAI and NVIDIA Unveil New Open-Weight AI Models for Global Inference Infrastructure
To ensure broad accessibility, OpenAI has partnered with numerous leading deployment platforms and hardware companies, including Azure, Hugging Face, vLLM, Ollama, llama.cpp, LM Studio, AWS, Fireworks, Together AI, Baseten, Databricks, Vercel, Cloudflare, OpenRouter, NVIDIA, AMD, Cerebras, Groq, and Microsoft. Microsoft, for instance, is bringing GPU-optimized versions of gpt-oss-20b to Windows devices, facilitating local inference for Windows developers.
