OpenAI publishes open weights for the first time since 2019
Two downloadable models under a permissive licence, released yesterday alongside a Claude upgrade. The company that argued against this has changed position.
On 5 August 2025 OpenAI released gpt-oss-120b and gpt-oss-20b: two reasoning models with published weights under the Apache 2.0 licence, its first open-weight release since GPT-2 in 2019. The same day, Anthropic released Claude Opus 4.1.
The Claude release is an incremental improvement. The OpenAI one is a change of position by the company that spent three years arguing that frontier weights should not be published.
Why this happened
Because the argument was being lost in the market rather than in debate. This blog has tracked it month by month: DeepSeek in January, Qwen in March and April, Llama throughout, Kimi K2 last month. Open-weight models went from curiosity to default option for a large part of the world’s developers.
A company whose models are not downloadable is absent from every deployment where downloading is the requirement. That is a commercial fact, and it eventually outweighs a position.
What it means for a Swiss SME
The smaller model is the interesting one. At twenty billion parameters it runs on hardware a company could actually own — the March position on the 32-billion model applies, one notch easier.
That matters for the same narrow group as before: firms with absolute confidentiality constraints. A law practice, a medical office, an engineering firm under a client NDA that forbids third-party processing. For everyone else, a hosted service under contract remains simpler and cheaper in total.
What has changed is that the option is now available from every major supplier rather than from challengers only. If you asked your vendor in March whether a self-hosted variant existed and were told no, the answer may have changed.
The pattern worth noting
Ten months ago the open and closed camps looked like different businesses with different customers. They are now the same companies offering both.
For a buyer that is unambiguously good. It means the choice between « hosted and convenient » and « self-run and private » is increasingly a configuration decision within one supplier relationship, rather than a choice of supplier. Fewer decisions that are expensive to reverse.