Kimi-K3 Beat Claude Fable at Frontend. The 2.8-Trillion-Parameter Weights Drop July 27

Cartoon illustration headlined “The open model took the crown,” showing three friendly computer-chip characters on a winners’ podium: an orange Kimi-K3 chip wearing a crown in first place, a grey Claude chip in second and a grey GPT chip in third, with a padlock and a “27 Jul” tag above and a “2.8T parameters” label in the corner. — *Kimi-K3 took #1 on the Frontend Code Arena over Claude Fable 5 and GPT-5.6 Sol, and the weights go open on 27 July.*

It took #1 on the human-judged Frontend Code Arena over Claude Fable 5, yet loses to GPT-5.6 Sol on every static benchmark. That gap, on the largest open model ever built, is the story.

Read for Free

Screenshot of a ray-traced black-hole simulator titled GARGANTUA, showing a glowing orange accretion disk warping around a dark event horizon, with an on-screen control panel listing render parameters such as observer distance, disk inclination and frame rate. — *The kind of output the arena was judging: a working, ray-traced black-hole simulator K3 generated as a single frontend task, from Moonshot’s own demo set. Source:* *Moonshot AI, Kimi-K3 technical blog*.

On 15 July 2026, the Frontend Code Arena crowned a new number one: Kimi-K3, from the Beijing lab Moonshot AI, at 1,679 points, ahead of Claude Fable 5 and GPT-5.6 Sol. The jump is the part that should stop you. The previous Moonshot model, K2.6, ranked 18th. K3 went from eighteenth to first in a single release and topped six of the arena’s seven frontend domains, taking second only in Gaming, behind Fable 5. Per Arena’s announcement, it is the first open-weight model to hold the top spot on that board. Moonshot ships the full weights on 27 July.

Max built a fully working Mac OS with Kimi-K3, which gives you a hint on how good this model is