Update: Self-hosting Kimi K3 needs datacenter GPUs, not your rig
The largest open-weight model ever is free to grab — but it wants ~1.4TB in 4-bit, a 64-GPU supernode, no llama.cpp path, and no cloud host yet.
Copy markdown
2.8 trillion params, all downloadable
Moonshot dropped Kimi K3's full weights on July 27 — 2.8 trillion parameters, which Bloomberg calls the largest open-weight model publicly available. It's a sparse mixture-of-experts that fires just 16 of 896 experts (~1.8%) per token, with a 1M-token context.
The hardware bill: ~1.4TB in 4-bit
Even at native MXFP4 (4-bit), the weights run ~1.4TB on disk. Moonshot's own guidance points at "supernode" rigs of 64+ accelerators; practical floors land near 8×H200 or 16×H100. This is a datacenter model, not a workstation one.
No llama.cpp, no Ollama, no GGUF yet
The only real serving path today is vLLM or SGLang, which ship Moonshot's KDA (Kimi Delta Attention) prefix-caching. There are no GGUF quants and no llama.cpp/Ollama/MLX support, so the "run it on my Mac" route simply doesn't exist at launch.
Data sovereignty is the actual point
No major cloud — Bedrock, Azure Foundry, Vertex — has integrated K3 yet, so it's Moonshot's China-hosted API or your own metal. Self-hosting is what lets teams keep code inside their own VPC instead of shipping it to an overseas endpoint.
A sovereignty play, not a capability upgrade
In Arena blind tests devs picked K3 ahead of leading U.S. models on front-end coding, but it still ranks below Claude Fable 5 and GPT-5.6 Sol overall. Self-hosting it is a control-and-cost move, not a jump past your current closed model — and a full technical report is still to come.