Kimi K3 Open Weights Are Here: How to Self-Host the 2.8T-Parameter Model
On July 27, 2026, Moonshot AI released the open weights for Kimi K3 -- a 2.8-trillion-parameter Mixture-of-Experts model that has been turning heads across the AI industry since its initial launch. The open-weight release means organizations can now deploy K3 on their own infrastructure, unlocking the same frontier-level reasoning capabilities without routing prompts through external APIs. This guide covers everything you need to know to get K3 running on your hardware, from downloading the weights to serving an OpenAI-compatible API.
What the Open-Weight Release Includes
Moonshot published the K3 weights on their Hugging Face organization under the Apache 2.0 license. The release includes:
MXFP4 quantized weights (~594 GB total download) -- the same format Moonshot uses internally, achieving near-lossless quality at 4-bit precision
