Kimi K3

Moonshot AI pushed open-weight scale to a new ceiling in July 2026 with Kimi K3, a 2.8 trillion parameter mixture-of-experts model that routes each token through 16 of 896 experts for roughly 104B active parameters. The design builds on Kimi Delta Attention, which Moonshot credits with sharply cheaper long-context inference, and a 401M parameter MoonViT-V2 encoder gives it native image understanding alongside text. Context stretches to 1,048,576 tokens, and a configurable thinking-effort control trades latency against deeper reasoning, which pays off in agentic coding and long-horizon knowledge work where K3 posts state-of-the-art open results. Weights ship in MXFP4 with MXFP8 activations for broad hardware compatibility, and recommended serving stacks are vLLM, SGLang, and TokenSpeed; a multi-GPU cluster is still required, so most teams will reach it through Moonshot's OpenAI and Anthropic compatible API endpoints instead. The Kimi K3 License, a modified MIT, keeps commercial use permissive for nearly all deployments.