gpt-oss-120b & gpt-oss-20b Model Card
OpenAI ● Covered by 2 sources
OpenAI just released two open-weight reasoning models, gpt-oss-120b and gpt-oss-20b, under Apache 2.0. That's a real license change for a company that's been closed-door for years.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI dropped something genuinely unexpected this week: two open-weight models, gpt-oss-120b and gpt-oss-20b, both released under the permissive Apache 2.0 license alongside a dedicated usage policy. For a company whose name has become a punchline about how little it actually opens up, this is a notable pivot, even if it's a narrow one.
Both models are built as reasoning systems, meaning they're designed to work through multi-step problems rather than just pattern-match to a quick answer. The naming convention makes the size difference obvious: 120b and 20b refer to parameter counts, giving developers a choice between a larger model with presumably stronger capabilities and a smaller one that's cheaper to run and easier to fit on modest hardware. That range matters. A lot of open-weight releases either aim for frontier bragging rights or for edge-device efficiency, rarely both in the same announcement.
What's missing from this initial rollout is the fine print that usually accompanies a model this size: benchmark numbers, training data specifics, context window details, or hard evidence of how it stacks up against competitors like Llama or Mistral's open offerings. OpenAI has published a model card, which is the right move for transparency, but the actual meat of that card, the safety evaluations, the intended use cases, the limitations, isn't summarized here. That gap will matter once developers start poking at these weights directly.
The Apache 2.0 license itself is the real headline. It's about as permissive as licensing gets, allowing commercial use, modification, and redistribution without the copyleft strings attached to something like GPL. Pairing that with OpenAI's own usage policy suggests they're trying to thread a needle: give the community real freedom while still drawing some lines around what they don't want their models doing. Whether that policy has teeth, or whether it's just words on a page, is the thing to watch once these models are actually in the wild and people start building on them.
My take — AI-written commentary, not fact-checked reporting
I'll believe OpenAI is serious about openness when I see the eval numbers and the training data disclosures, not just a license file. Apache 2.0 is genuinely generous, and credit where it's due, but a model card without benchmarks is a press release wearing a lab coat. Watch what the community actually builds with the 20b variant on consumer hardware before deciding this changes anything.
Read more about this at: OpenAI