GitHub radar
Kimi-K3 MLX Port: Moonshot AI's 2.78T Model for Apple Silicon
A community developer has ported Moonshot AI's Kimi-K3 — a 2.78-trillion-parameter multimodal model with a 1M-token context window — to Apple's MLX framework, dissecting every architectural layer including the novel Kimi Delta Attention and Attention Residuals mechanisms.
PipeNetwork ported Moonshot AI's Kimi-K3 to Apple's MLX framework (Apple Silicon). Kimi-K3 is a 2.78-trillion-parameter multimodal MoE model — 104 billion parameters active per token — built by the Chinese AI company Moonshot AI with a 1M-token context window. The port covers the full text tower (Kimi Delta Attention + Attention Residuals), the 3D-capable video vision tower MoonViT, and a multimodal wrapper, each validated with numerical parity tests against the original PyTorch code.
Why a vibe-coder should care
Running Kimi-K3 requires at least 870 GB of RAM — more than the 512 GB ceiling of the most powerful Mac ever made — so this port cannot run on any single machine today. Its real value is in making the model's internals fully legible on Apple's ML stack: when the architecture is this well documented, hosted API services and cloud deployments of K3 arrive faster.
▌ More finds