Deltafin is a research project enabling the Kimi K3 LLM, a 2.8T-parameter Mixture-of-Experts model, to run on a single Apple Silicon Mac. It streams experts on demand over HTTP into a local disk cache, leveraging fused NEON kernels, Metal/MPS compute, and an OpenAI-compatible API server.

Installation involves three commands: setting up the environment, building a fused MXFP4 kernel, and downloading the model. Users can choose a full ~1.7 TB install for faster inference or a streaming mode requiring ~215 GB, which downloads experts as needed. The project highlights that while inference is slow (around 16 seconds per token on an M1 Max), it offers exact, reproducible decoding.

Deltafin provides an OpenAI-compatible API server, allowing local chat and coding agents to interact with the model. Users can switch to a full install from streaming mode with a single command. The project emphasizes the trade-offs between disk space, download time, and inference speed, noting that newer Apple Silicon chips and more RAM will automatically improve performance.