Kimi K3 is a 2.8T open-weight MoE, but the weights are not out yet
Moonshot AI announced Kimi K3, a 2.8-trillion-parameter MoE model, on July 16. Open weights are promised by July 27. Until then it is API-only, and running it yourself will need serious cluster hardware.
Moonshot AI announced Kimi K3 on July 16, describing it as their “most capable model to date” at 2.8 trillion parameters. The open-weight release is promised by July 27, 2026, but as of today the weights are not on Hugging Face. API access is live now at $3 per million input tokens and $15 per million output tokens, which puts it in the same bracket as Anthropic’s Sonnet series and makes it the most expensive model from a Chinese lab to date, per Simon Willison’s write-up.
The architecture is a sparse MoE: 896 experts total, 16 active per token, built on a new attention mechanism Moonshot calls Kimi Delta Attention. Self-reported benchmarks put K3 mostly ahead of Claude Opus 4.8 and GPT-5.5, and behind Claude Fable 5 and GPT-5.6 Sol. Willison ran his pelican SVG benchmark against it and noted something worth flagging: K3 currently only has one reasoning effort level, “max,” and it shows. A simple SVG generation prompt burned 13,241 reasoning tokens and cost him 25 cents. That is the default behavior until Moonshot ships lower-effort modes.
The hardware reality for self-hosting is brutal. Moonshot’s own deployment guidance points at 64 or more accelerators for practical serving. This is not a model you are going to run on a Mac Studio or even a small GPU cluster. When the weights do drop, the story for most people will be quantized builds via vLLM or llama.cpp, and even then you are looking at multi-GPU territory. Watch the Moonshot HuggingFace org on or after July 27 for the actual files, the license, and the technical report. Treat the date as a target, not a guarantee.