Moonshot Ships Kimi K3 — 2.8 Trillion Parameters, First Open 3T-Class Model, and It Beats Fable 5 on Arena's Frontend Code Board

VentureBeat·high signal

Moonshot AI released Kimi K3, a sparse MoE activating just 16 of 896 experts per token (~1.8% of the pool) with a 1M-token context window and native vision, introducing two new architectural pieces — Kimi Delta Attention and Attention Residuals. It ranked first in Arena's Frontend Code evaluation at 1,679 points, ahead of Claude Fable 5 in blind developer testing, while trailing Fable 5 and GPT-5.6 Sol on overall capability. Full weights land by July 27 under a Modified-MIT-style license at $0.30/M cache-hit input, $3/M cache-miss, $15/M output — meaning a frontier-adjacent coding model becomes self-hostable in under a week.