Language

Kimi K3 (MoE, multimodal)

Moonshot AI Kimi K3 — 2.8T total parameters, 104B active, 896 routed experts with 16 selected per token and 2 shared. Kimi Delta Attention plus gated MLA across 93 layers, 1M context, 160K vocabulary, MXFP4 weights and MXFP8 activations from quantization-aware training. Text, image, and video via the MoonViT-V2 encoder. `UD-IQ1_S` (594GB) is the smallest quant; `UD-Q2_K_XL` (861GB) is the size/quality balance point. Both exceed any single machine, so whole-model serving means a pipeline cluster; a lone host runs it as distributed expert extraction instead.

Kimi K3 LicenseSource verified

Specification

Model ID

kimi-k3

Family

kimi-k3

Modality

Language

Parameters

2.8T total / 104B active (MoE)

Context

1,048,576 tokens

Quantization

UD-IQ1_S

Weights

553.2 GB

Minimum RAM

610 GB

Source

Weights and provenance.

Repository

unsloth/Kimi-K3-GGUF

Registry license

Kimi K3 License

Access

Open — weights fetch without accepting additional terms.

The licence above is the registry's. The source repository either states nothing, states the catch-all “other”, or carries a licence link and base model that its own tag contradicts — a re-quantisation inherits the terms of the weights it derives from. The registry's licence is the one to rely on.

Serve it

# Pull the weights onto a node
tenzro model download kimi-k3

# Serve it
tenzro model serve kimi-k3

# Or query the registry entry over JSON-RPC
curl https://rpc.tenzro.xyz \
  -X POST -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,
       "method":"tenzro_getModel",
       "params":[{"model_id":"kimi-k3"}]}'

← All models