open-weights

By late September 2026, models whose weights can be downloaded carried about 60% of the tokens routed through OpenRouter but only 26% of the spend, a split published by Peter Walker in OnlyCFO on 22 October 2026. Mozilla dated the crossover to mid-2026, with the seven highest-volume models all open, while noting that closed providers still led by number of requests.

The term is narrow: the weights are downloadable and runnable, the training pipeline is not, and the OSI definition of open source AI also demands training code and data documentation. Two thirds of those tokens come from Chinese labs, though hosting is chosen separately from the model, and American providers serve 44% of them.

What the source notes keep circling is reversibility. SFEIR's reading of Kimi K3 (Moonshot AI, 16 July 2026) states it directly: a proprietary model is consumed, an open-weights model is recovered as an option, to be run or ported. In the Mistral and Microsoft agreement of 21 July 2026, the open-weights character of Mistral's models is the single element that makes the sovereignty claim portable.

Dean W. Ball reads the same trend inversely. In a thread of 17 July 2026 he argues that publishing weights deters AI capex and that a generalised open-weights world ends in state-run AI as public infrastructure, which he calls dystopian. Mozilla supplies a different friction: 53% of teams working in open reach production against 63% in closed, and the gap widens with company size.