Welcome GPT OSS, the new open-source model family from OpenAI!

OpenAI has released a new open-weights model family called GPT OSS, featuring a 117B and a 21B parameter version. These models are Mixture-of-Experts (MoEs) that use a 4-bit quantization scheme, allowing the large model to run on a single H100 GPU and the smaller one on consumer hardware. Licensed under Apache 2.0, the models support powerful reasoning, tool use, and are designed for broad accessibility in the open-source community. The architecture includes features like token-choice MoE, RoPE with a 128K context, and alternating attention layers, and can be used locally with tools like transformers and vLLM.