gpt-oss-120b reaches near-parity with o4-mini on a single 80GB GPU, under Apache 2.0
OpenAI released two open-weight models: 117B parameters with 5.1B active fitting one 80GB card, and a 21B model with 3.6B active that runs on edge devices with 16GB. Both Apache 2.0, both natively quantised in MXFP4, weights on Hugging Face.
The single-GPU claim is the one that changes who can participate. A 117B mixture-of-experts with 5.1B active parameters fitting on one H100 or MI300X means near-o4-mini reasoning is available to any lab, university or company that can rent one card by the hour.
Apache 2.0 is the other half and arguably the larger one. Most large open-weight releases carry bespoke community licences with usage thresholds and field-of-use restrictions that require legal review before integration. Apache 2.0 removes that step entirely.
The 20B model reaching o3-mini-level results in 16GB is the edge story. That is a laptop, and it is the same argument being made further down the size curve — capability is moving to hardware that was never expected to hold it.
One caveat the model card states plainly: both models were trained on OpenAI's harmony response format and will not work correctly without it. That is a real integration constraint, not a footnote.
OpenAI — Introducing gpt-oss → · OpenAI — gpt-oss-120b and gpt-oss-20b model card → · Hugging Face — openai/gpt-oss-120b →