Moonshot releases Kimi K3 open-weights model with agentic tooling and near-frontier performance

Moonshot AI shipped Kimi K3 as an open-weights model, validated by multiple independent evaluations to surpass Opus 4.8 on the Artificial Analysis Intelligence Index, narrowing the gap between leading proprietary and open models to 4 points.

The model is a 2.8T-parameter mixture-of-experts architecture with 104B active parameters, 896 experts (16 active per token), 1M-token context, and native visual understanding, according to Moonshot’s technical report and companion posts.

Moonshot also open-sourced supporting infrastructure: FlashKDA (Kimi Delta Attention kernels), MoonEP (MoE communication library), and AgentENV (distributed agent environment), positioning K3 as a full recipe for large-scale agentic post-training and serving.

Licensing is described as “open weights,” not permissive open-source; large hosting providers over $20M/year revenue require a separate agreement, and products above 100M monthly active users or $20M/month revenue must display “Kimi K3” in the UI, per multiple observers citing the terms.

Kimi K3 was immediately available across vLLM, Baseten, Modal, Fireworks, Nebius, Together, DigitalOcean, Cursor, Cognition/Devin, Ollama Cloud, and Dell Enterprise Hub, underscoring open-weight launches as supply-chain events.

Early evaluations on Agent Arena report Kimi K3 Max ranks first among open-weight models with a +9.75% net improvement across multiple agentic signals, including confirmed success rates.