Unsloth has compressed Moonshot AI’s Kimi K3 model into a 594GB dynamic 1-bit format, a 62% reduction from its original 1.56TB that allows the model to run on a Mac Studio with 128GB of RAM. The updated version retains approximately 78.9% of the model’s baseline accuracy.

The update follows the open-weight release of the 2.8T parameter model, which typically requires massive server clusters and dozens of GPUs to operate. Unsloth plans to push the weight size further to 512GB, lowering hardware barriers for developers running the frontier model independently.