An open source model just beat Claude at frontend code, and you can use it for free right now. Moonshot AI dropped Kimi K3, and at 2.8 trillion parameters it is the biggest open model ever built.
Number one at frontend code
On Arena.AI's frontend code arena, where real humans vote head to head on which output they prefer, Kimi K3 sits at number one with a 1,679 Elo. Claude Fable 5 and GPT 5.6 Sol both land below it. For anyone shipping UI, that is the benchmark that matters, because it measures what people actually pick, not what a script scores.
Where it ranks on everything else
Frontend is not the whole story, so here is the honest version. On the Real Work leaderboard (GDPval-AA v2 Elo), Kimi K3 comes in third at 1,668, behind Fable 5 at 1,760 and GPT 5.6 Sol at 1,748, but still ahead of Claude Opus 4.8.
Then it flips back to first on deep web research. On BrowseComp it scores 91.2, a new state of the art, edging out GPT 5.6 Sol at 90.4 and Fable 5 at 88.0. So Kimi K3 is a genuine frontier model that happens to be strongest exactly where most people build: the front end and research.
It designed its own chip
Here is the part that sounds made up. Moonshot let Kimi K3 run alone for 48 hours as a long horizon agent, and it designed a working chip. Four square millimeters, 100 MHz timing closure, roughly 8,700 tokens per second in simulation, and the chip runs a nano version of K3 itself. Full pipeline, from architecture all the way through verification, with no human in the loop. That is what long horizon agents look like now.
How to use Kimi K3 for free
The best part is the price. Go to kimi.com, sign in with a Google account, and you get access with no credit card. You get the K3 model with a 1 million token context window at zero cost. And on July 27th, Moonshot is releasing the full weights, so you will be able to download the entire model and run it yourself.
An open source model beating closed frontier labs on the benchmarks people care about, for free, is the story of 2026 in one release.