For three years, one line reassured every executive nervous about open-weight AI: the free models were good, but they trailed the closed frontier by about six months. Download something today and you were running last winter's best. This week that comfortable lag looked, suddenly, close to gone. Two releases from Chinese labs landed within days of each other, and both claimed to sit level with the most expensive systems money can rent.
The louder of the two was Kimi K3, from Beijing's Moonshot AI. At 2.8 trillion parameters it is, the company says, the largest open model ever built, roughly 75 percent bigger than DeepSeek's flagship. Numbers that large invite eye-rolling, but the benchmarks are what caught attention. On private evaluations run by the analytics firm Artificial Analysis, K3 placed third on a test of real-world professional tasks and second on a measure of long, multi-step knowledge work, trading places with the top models from OpenAI and Anthropic rather than trailing them. In one demonstration it reportedly ran for 48 hours unattended and designed a small working chip to run a miniature version of itself.
The quieter release may matter more to anyone who writes code. Z.ai, the lab formerly known as Zhipu, put out GLM-5.1, a 754-billion-parameter model under the permissive MIT licence, and by its own account it took the top spot on SWE-Bench Pro, one of the hardest coding tests, edging ahead of the leading models from OpenAI, Anthropic and Google. The weights are free to download. A widely shared line summed up the week's mood: open source is no longer six months behind. Read that again, the argument went, and think about what it means.
It is worth reading the claims sceptically. These are, for now, mostly self-reported scores, and benchmark numbers have a way of shrinking once outside researchers get their hands on the weights. The precise version numbers of the Western models being beaten differ from one write-up to the next, a reminder that the leaderboard moves faster than anyone can check it. The honest summary is not that open models have won, but that a distance once measured in seasons is now measured in weeks, if that.
There is a catch hiding in the word "open," though. A 2.8-trillion-parameter model is open in the sense that its weights are public, and no one can revoke your access or read your prompts. It is not open in the sense of running on your laptop; inference at that scale still needs a rack of expensive accelerators. So the real shift is not that powerful AI suddenly became cheap to run. It is that the frontier is no longer something only a handful of American companies can meter out to you through an API. For anyone building on top of these systems, and for the governments watching which models quietly become the world's default, that is the number that counts.