Overview
- Kimi publicly rolled out K3 over July 17–18 and opened trials on kimi.com, the Kimi app and APIs while announcing full weights for release by July 27.
- Independent tests and leaderboards placed K3 first on several web and front‑end engineering benchmarks but Kimi concedes the model still trails the strongest closed systems on overall metrics.
- K3 introduces Kimi Delta Attention, Attention Residuals and a higher‑sparsity mixture‑of‑experts design that Kimi says yields roughly 2.5× scaling efficiency versus its prior K2 models.
- Hands‑on demos show strong multimodal and code‑generation results, including one‑shot 3D/HTML game generation and long agent runs that Kimi used to illustrate complex workflows such as a claimed 48‑hour chip design run.
- Kimi set higher prices than its prior tiers but below the top closed APIs, has limited free trial quotas, and heavy demand caused temporary access limits during early testing which could affect developer uptake in the short term.