OpenAI opened the weights, and the licence is the news
Near-o4-mini reasoning on a single 80GB card matters. Apache 2.0 matters more, because it removes the step that actually blocks adoption.
gpt-oss-120b runs on one 80GB GPU with 5.1B active parameters and reaches near-parity with o4-mini on core reasoning. The 20B sibling hits o3-mini-level results in 16GB. Both Apache 2.0, both natively quantised in MXFP4.
Why the licence outweighs the benchmark
Most large open-weight releases arrive under bespoke community licences carrying usage thresholds, field-of-use restrictions and acceptable-use policies. Every one of those requires legal review before a single line of integration code gets written.
Apache 2.0 deletes that meeting. For an enterprise that is the difference between a two-week procurement conversation and a git clone, and it is a bigger practical unlock than several points on a benchmark.
The single-GPU threshold
A 117B mixture-of-experts fitting one H100 or MI300X means frontier-adjacent reasoning is available to anyone who can rent a card by the hour. That is universities, small labs, and companies that cannot send data to a third party.
And the pressure continues downward. A 2.6B model doing agentic work on a Raspberry Pi is not competing with any of this on reasoning — it is competing with having nothing on the device, which is a comparison it wins outright.
The part that does not travel
Capability parity is now measured in months. The safety apparatus does not narrow with it.
Staged rollout, usage monitoring, revocation, the ability to patch every deployment at once — none of that ships with a weights file. An open release is final. Whatever properties it has on release day are the properties it has permanently, in every fork, forever.
That is not an argument against opening weights. It is an argument that the safety question for open models is a different question, and it is currently being answered by whoever downloads the file.
OpenAI — Introducing gpt-oss → · OpenAI — gpt-oss-120b and gpt-oss-20b model card → · TechCrunch — Open-weight AI models are catching up to the frontier. The safety gap remains →