OpenAI gpt-oss (120b / 20b)

Both checkpoints are downloadable from Hugging Face and GitHub and were built to run locally, from a laptop upward.

Apache 2.0 on the weights, with a usage policy sitting alongside that does not withdraw or condition the software grant; commercial use, modification, and redistribution are permitted.

Designed for local and independent deployment with hardware-optimized builds for consumer GPUs and support across vLLM, Ollama, LM Studio, and major clouds.

The dataset is described at a high level along with architecture, tokenizer, post-training approach, and safety evaluation, but no reproducible training corpus or recipe is published.

Limited to moderate: the training data and full training pipeline are not released, so an equivalent system can be approximated but not rebuilt.

OpenAI's openai/gpt-oss repository accepts a small number of merged outside pull requests to the reference implementations and harmony tooling, but the model, its post-training, and the release decision are entirely OpenAI's.