Open Weights Just Caught Up to the Frontier and Mostly Nobody Noticed
Kimi K3.1 Pro, Mistral Large 4, Tencent Hy4, all open and all genuinely good. The gap the labs sold us for two years is now a rounding error, and the people who should care are busy arguing about chatbots.
The story the feeds tell is always the next chatbot, but the story that will matter in a year is that open weights caught up and the people who should care were looking the other way. This month Kimi K3.1 Pro went open as the leading open model, Mistral shipped Large 4, the trillion-parameter Le Chonk, under an open licence, and Tencent released Hy4 with native visual reasoning and computer-use.
For two years the closed labs sold a gap, the idea that what they kept behind the API was years ahead of anything you could run or modify. That gap is now a rounding error on the benchmarks that count, and more importantly it is an aperture, open models mean you can fine-tune, self-host, and audit, which a hosted API simply will not let you do.
The practical upside is sovereignty, the word the releases keep using, and it is not vanity. A company that can run the model inside its own walls stops sending its data to a frontier provider and stops reforecasting when that provider cuts a price. For regulated industries and for anyone outside the US and EU cloud orbit, that independence is the feature.
The catch is the one open source always carries, you now own the operations. Running a trillion-parameter model is not free and not trivial, and the licence fine print, revenue thresholds, regional exclusions, actually matters once you ship. The cheap-and-open era is real, but it trades a subscription for a sysadmin.
What I find striking is the silence. A genuinely open model beating closed ones on agent benchmarks should have been the headline, and instead it was a line under a chatbot launch. The catch-up happened; the audience has not caught up to it yet.