Kimi K3 Open Weights Arrive Sunday: Self-Hosting Cuts China Data Risk the API Never Can

July 25, 2026

Moonshot AI will publish full model weights for Kimi K3 — billed as the largest open-weight AI model ever released at 2.8 trillion parameters — on Hugging Face at July 27, 2026, 00:00 UTC (July 26, 8:00 PM ET).

Teams face a choice: Moonshot hosted API vs download-and-self-host. Self-hosting means prompts need not touch servers subject to China’s National Intelligence Law / Cybersecurity Law / Data Security Law at the inference layer.

Architecture notes

K3 is a sparse MoE with 896 experts, ~16 active per token (~50B-dense-equivalent compute). Innovations cited: Kimi Delta Attention (KDA) for 1M-token context (up to 6.3x faster long-context decoding claimed), Stable LatentMoE, Attention Residuals. MXFP4 weights ~1.4 TB fast memory / ~594 GB download per advance tracking; Moonshot recommends 64+ accelerators (NVIDIA Blackwell / AMD MI400 for native MXFP4).

Benchmarks and caveats

Arena.ai Frontend Code Arena: first place (1,679 Elo). Artificial Analysis Intelligence Index v4.1: 57.1 (4th; behind Fable 5 and GPT-5.6 Sol, ahead of Opus 4.8). Speed/latency and ~51% hallucination rate on AA-Omniscience noted as weaknesses. K3 license not yet published as of July 25 — verify repository license before commercial use.

Geopolitics

White House OSTP (July 22) accused Moonshot of distilling Anthropic Fable; Bessent threatened sanctions. FT (July 21): China Commerce Ministry consulting firms on restricting foreign access to AI weights — not finalized. Moonshot says July 27 release proceeds on schedule. Access: Hugging Face (weights), platform.kimi.ai model id kimi-k3 (API), kimi.com (web).