Changes from previous version
Successor to Kimi K2.6 with a new architecture (Kimi Delta Attention + Attention Residuals), roughly 2.8T parameters, native multimodal vision/video, always-on reasoning, and a 1M context window (up from K2.6's 256k). Moves from open-weight K2.x agent pricing to Sonnet-tier API rates ($3/$15 with $0.30 cache hits).
Release Summary
Moonshot AI's most capable Kimi model to date: 2.8 trillion parameters with Kimi Delta Attention (hybrid linear attention) and Attention Residuals, native visual understanding (text, image, and video input), and a flat 1M-token context window. Always-on thinking via reasoning_effort (currently max only). Aimed at long-horizon software engineering, knowledge work, and deep reasoning. API model ID kimi-k3 on the Moonshot OpenAI-compatible endpoint. Flat pay-as-you-go pricing with automatic context caching: $0.30/M cache-hit input, $3/M cache-miss input, and $15/M output (same list rates as Anthropic's Claude Sonnet series, close to GPT Terra 5.6 at $2.5/$15). Launch reporting places it competitive with current frontier models and behind Claude Fable 5 and GPT Sol 5.6. Real cost still depends on reasoning-token efficiency versus leaner models like Sol. Available on kimi.com, Kimi Code, and the Kimi API; public weight downloads were not yet published at launch.
Timeline
Kimi K3 released
Moonshot AI launches Kimi K3 on the Kimi API Platform and kimi.com with 2.8T parameters, 1M context, native vision, and $0.30/$3/$15 per-million-token pricing.