Kimi K3 is an open-weight, native multimodal agentic model from Moonshot AI.
118
Kimi K3 is an open-weight, native multimodal agentic model from Moonshot AI. It is a 2.8-trillion-parameter Mixture-of-Experts model (16 of 896 experts active per token, 104B active) built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision via MoonViT-V2 and a 1-million-token context window. Weights are quantization-aware-trained MXFP4 with MXFP8 activations. Kimi K3 always has thinking enabled and returns reasoning_content; thinking effort is configured with reasoning_effort (low, high, max). Tool calling uses the kimi_k3 parser. Both the code repository and the model weights are released under the Kimi K3 License.
| Vendor | AMD |
| License | Kimi K3 License, MIT |
| Source | https://github.com/amd-enterprise-ai/aim-build |
| Revision | 31cc841d99d82ca2236c15d753f7e4166007bfb1 |
| Built | 2026-10-22T08:15:55Z |
Content type
Image
Digest
sha256:1955b668a…
Size
16.2 GB
Last updated
16 days ago
docker pull amdenterpriseai/aim-moonshotai-kimi-k3:2026.11.0-rc1