Kimi K3: The Largest Open-Weight Frontier Model Ships (2.8T MoE, 1M Context, Weights July 27)

TL;DR: Moonshot AI's Kimi K3 — a 2.8T-parameter MoE with 1M context — debuted July 16 at #1 on the Frontend Code Arena leaderboard; weights release July 27 under Modified MIT.

Kimi K3 is a 2.8-trillion-parameter Mixture-of-Experts model with 896 total experts and 16 active per token. Two new architectural components drive it: Kimi Delta Attention (KDA), a hybrid linear attention mechanism that reduces quadratic attention cost on long sequences, and Attention Residuals (AttnRes), which Moonshot says improves long-context reasoning quality. Context window: 1M tokens. Native vision support. Priced at $3/$15 per million tokens — same as Claude Sonnet 5.

On independent evaluation it ranks fourth among all frontier models — behind Claude Fable 5 and GPT-5.6 Sol, ahead of Claude Opus 4.8 — and sits first in the Frontend Code Arena with 1,679 points. Full weights land July 27 under a Modified MIT license.

Why it matters: When weights release July 27, Kimi K3 will be the largest open-weight model available to self-host. If it holds its leaderboard position in practice, it opens a real path to running frontier-grade coding assistance without API dependency — worth watching closely.

Sources: Kimi K3 — MarkTechPost · Moonshot AI Kimi K3