For two years the case for open-weight models rested on price and control. You could run them on your own hardware, inspect their internals, and skip the API bill. This month a third argument arrived, and it is the one enterprise risk teams cannot ignore: a model you have downloaded cannot be taken away from you by a government letter.
The capability case was already getting hard to dismiss. When Z.ai's GLM-5.2 landed on 13 June, Vercel's chief executive Guillermo Rauch said he was "genuinely impressed, almost shocked" by its coding, and his company wired it into its AI Gateway within three days. The numbers explain the reaction. On FrontierSWE, a demanding coding benchmark, GLM-5.2 trails Claude Opus 4.8, the leading closed model, by a single percentage point. Its Terminal-Bench score jumped from 63.5 to 81.0 in one generation, and its context window grew fivefold to a million tokens. The weights are on Hugging Face for anyone to run locally.
A one-point gap to the best closed model is the kind of number that makes premium API pricing look optional. But pricing is the old argument. The new one was supplied, unintentionally, by the US government. When Washington ordered Anthropic's Fable 5 pulled offline worldwide, it converted an abstract worry about data sovereignty into a live outage. The most capable public model in history went dark within hours of a directive, and there was nothing customers could do about it.
China's MiniMax moved immediately to make the point explicit, positioning its M3 model for teams that lost Fable 5 access and noting what no closed provider can match: open weights cannot be recalled by any government. A US-hosted, closed model can be globally disabled in an afternoon. A model whose weights already sit on servers in Frankfurt, Seoul or Singapore cannot be. M3 joins GLM-5.2, Kimi K2.7-Code and Meta's Llama 4 as the self-hostable options now drawing serious enterprise evaluation.
None of this means the gap has closed entirely. On SWE-Bench Verified, the best self-hostable models still sit roughly six to eight points behind the strongest available closed alternative, a real difference for hard engineering work. The honest framing is a trade. Closed models still lead on the frontier; open ones now offer a capability that is merely close, plus immunity from being switched off. Before this month, that immunity was a talking point. The Fable 5 ban turned it into a line item, and notably most of the credible challengers carrying it are Chinese. The irony is sharp: an export control meant to keep the best AI inside American hands has become the best marketing the open-weight world could ask for.