Kimi K3 publishes specs, pricing, and benchmarks with weights due 2026-07-27

Sources: Moonshot blog, Kimi K3 API pricing, Simon Willison, Artificial Analysis, HN discussion
Summary: Moonshot AI published a Kimi K3 announcement on 2026-07-16, adding the model detail that was absent when K3 went live in the product the same day. K3 is a 2.8T-parameter Mixture-of-Experts model (Stable LatentMoE with 16 of 896 experts active per token) built on Kimi Delta Attention, with a 1M-token context window and native multimodal input. The API model id is kimi-k3, priced at $0.30 per 1M cache-hit input tokens, $3.00 per 1M cache-miss input, and $15.00 per 1M output. Full weights are promised by 2026-07-27, with a technical report to follow.
Comments: HN commenters note the $3.00/$15.00 per 1M pricing (cache $0.30) is high for a Chinese open-weight model, and that Moonshot's own figures rank K3's overall intelligence behind Claude Fable 5 and GPT-5.6 Sol.
Why it matters: A 2.8T open-weight model with a scheduled weight release sets a new size point for self-hostable frontier models, though the benchmarks are vendor-reported and unverified until the report and weights land.
Follow-up: Watch for the 2026-07-27 weight release and license, the technical report, and independent benchmark reproduction.