Moonshot’s Kimi K3 now runs locally at 594 GB with 78.9% top-1 accuracy

danielhanchen · x · 2026-07-29

Moonshot’s Kimi K3 is now runnable locally through Unsloth’s quantized builds.

The model is described as a 2.8T-parameter open-weight model with 104B active parameters, native vision, and a 1M-token context window.
Unsloth reports a dynamic 1-bit build at 594 GB, down from 1.56 TB, while retaining about 78.9% top-1 accuracy.
A dynamic 2-bit build reaches about 90% accuracy at 861.3 GB, and the model can run on setups such as a Mac Studio with 128 GB RAM or an NVIDIA DGX Station-class machine.
The guide also includes hardware requirements, quantization comparisons, and notes on running the model via Unsloth Studio or llama.cpp.

More from coding & agent