gpt-oss: OpenAI releases open-weight models under Apache 2.0
OpenAI released gpt-oss-120b and gpt-oss-20b, mixture-of-experts reasoning models quantised with MXFP4 and licensed Apache 2.0; the large one fits on one H100 and the small one in 16GB.
What happened
OpenAI released GPT OSS, open-weight models designed for reasoning, agentic tasks and developer use, with day-one Hugging Face support (Transformers with Flash Attention 3, llama.cpp, vLLM, transformers serve and cloud partners).
The technical details
From the Hugging Face post:
- two models: gpt-oss-120b (117B parameters) and gpt-oss-20b (21B)
- both mixture-of-experts, with few active parameters per token (the post’s overview lists 5.1B and 3.6B)
- weights use a 4-bit MXFP4 quantisation scheme
- the large model fits on a single H100, and the small one runs in 16GB
- licensed Apache 2.0 with a minimal usage policy
- support for system and developer messages, and tool use
Why it matters
The post quotes OpenAI framing this as a meaningful step in its commitment to the open-source ecosystem. Technically, shipping weights already in a 4-bit format shows how standard quantisation (my March 2026 study notes) has become for open models.
#llm#open-weights#mixture-of-experts
Sources
This is a summary written from the linked public sources. Figures and numbers are as reported there.