Transmission // 10
The weights came from Hangzhou
2 min
Four of the five leading open-weight models come from Chinese labs. "Open source means second best" stopped being true, and the licences deserve a closer look than the benchmarks.
The assumption that the strongest models necessarily sit behind a corporate API has quietly collapsed. Top open-weight models now cluster around 80% on SWE-bench Verified and 90% on GPQA Diamond — the same conversation as the closed frontier, not a tier below it.
The geography is the part nobody planned. Four of the five leaders are Chinese: DeepSeek, Moonshot, Zhipu, Alibaba’s Qwen. DeepSeek’s reasoning line ships under a plain MIT licence. Qwen 3 spans 0.5B to 235B parameters — phone to datacentre from one family, one tokeniser, one prompt format.
“Open” is doing a lot of work in that sentence
Meta’s Llama 4 is the most-searched Western open model and the Open Source Initiative does not classify it as open source, because the community licence carries commercial-scale limits and field-of-use restrictions. That is not a technicality. It is the difference between downloadable and yours.
The licence spectrum, in the order that matters to anyone building:
- MIT / Apache 2.0 — actually open. Fork it, sell it, embed it, no conversation.
- Community licences with scale caps — fine until you succeed.
- Open weights, closed data, closed recipe — reproducible in the sense that a cake is reproducible from a slice.
Almost nothing on the leaderboard is open in the third sense, including the models people call the most open. Weights are an artefact, not a method.
Why this matters more than the ranking
A closed model is a dependency with a phone number. It can be deprecated, repriced, safety-tuned into a different personality, or made unavailable in your jurisdiction — and none of those events are yours to schedule. Every one of those has happened to somebody’s production system in the last two years.
An open-weight model is a file. Files do not get deprecated. They get old, which is a completely different failure mode and one you control the timing of.
The gap between the best open model and the best closed model is now smaller than the gap between a model you can pin and a model somebody else can change underneath you.
Pin the file. Ship on top of something that cannot be recalled.