A model with no open weights and no technical report has climbed to first place in frontend code benchmarks in just a couple of days, beating Fable 5 and Opus 4.8.
Benchmarks
That's Kimi K3 from Chinese Moonshot AI — one million token context window, focused on code and agentic tasks. It's sitting at number one in Frontend Code Arena with 1679 points; Fable 5 has 1631, GPT-5.6 Sol xHigh has 1618. In Artificial Analysis's general Intelligence Index, K3 ranks third with a score of 57, behind only Fable 5 (60) and GPT-5.6 Sol (59), but already ahead of Opus 4.8 (56).
Parameters and codename
Before the official release the model was tested under the codename Kivine, and one prompt for a static scene produced an animated Star Wars-style Death Star attack sequence. Moonshot is claiming 2.8 trillion parameters — more than DeepSeek V4-Pro (1.6T) — and says this is the ninth time in a year that Kimi has set a size record among open models. Pre-release leaks had put it at 2.5T, off by 300 billion.
Weights and pricing
No weights yet. Moonshot says they're coming "soon," and the license is supposedly the same Modified MIT as K2, but that's unconfirmed. API pricing is $0.94 per task — more expensive than DeepSeek, but several times cheaper than Fable 5 at $2.75.
All these numbers are from Moonshot itself. I'd hold the applause until independent benchmarks are in.