kimi-k3-mlx
An MLX port of Moonshot AI's 2.78-trillion-parameter Kimi-K3 multimodal mixture-of-experts model for Apple Silicon, with a streaming weight converter, expert pruning and per-language expert-overlap analysis, though no single Mac can currently run it.
This project provides a stock mlx-lm-compatible model definition for Kimi-K3's text tower, a matching vision tower and projector, and an mlx-vlm wrapper joining the two, so the model can be used with the standard MLX ecosystem on Apple Silicon. Because the source model is roughly 5.6 TB in bf16, a custom streaming converter is required in place of the ordinary mlx_lm convert, which would otherwise need to materialize the whole model at once.
The architecture combines 896 routed experts, top-16 per token, processed in a compressed latent space, a mix of Kimi Delta Attention and gated multi-head latent attention layers, and a Stable LatentMoE and Attention Residuals design from the original model. The README states plainly that nothing in the repository runs on a single Mac today: the smallest deployable tier is roughly 870 GB against a 512 GB ceiling on the largest existing Apple Silicon machine.