OpenAI has unveiled two new open‑weight language models, gpt‑oss‑120B and gpt‑oss‑20B, available free for developers, researchers, and companies worldwide.
The smaller 20B model runs on a standard laptop with 16 GB RAM, while the 120B model demands an 80 GB GPU for optimal performance.
Unique to the release is open‑weight access, developers can inspect and fine‑tune internal parameters, but the training dataset remains undisclosed, so it doesn’t meet full open‑source criteria.
The 120B model matches or exceeds the capacity of OpenAI’s proprietary o4‑mini on reasoning benchmarks, while the smaller version rivals o3‑mini, including coding and competition math tasks. These text‑only models support chain‑of‑thought reasoning, web browsing, code execution, and agentic workflows under OpenAI’s Responses API.
The release follows global attention on China’s DeepSeek-V2 and V3 open‑weight systems and Meta’s LLaMA series, pressuring OpenAI to offer its own transparent model stack.
Amazon Web Services is integrating GPT‑OSS into Bedrock and SageMaker JumpStart, enabling developers to deploy models that are up to three times more cost-efficient than competitors like Gemini or DeepSeek‑R1.
OpenAI delayed the launch twice to conduct rigorous safety testing against misuse risks. Internal adversarial testing confirmed GPT‑OSS does not reach “High capability” in biological or cyberattack domains, even after simulating fine‑tuned misuse. OpenAI confirms this release is the first open‑weight model since GPT‑2, launched in 2019
OpenAI users can experiment with the new models on its open model playground. It also provided guides for those interested in using the models through different ecosystem providers or fine-tuning them for a specific task.
