Someone Is Running Kimi K3's 2.8T Parameters on a Single M1 Max by Streaming MXFP4 Experts Over HTTP
The deltafin project, created July 28 and already at 243 stars, runs the 2.8T-parameter Kimi K3 MoE on one Apple Silicon Mac by streaming MXFP4 expert weights on demand over HTTP into a local disk cache rather than resident memory, with fused NEON kernels, Metal/MPS compute, exact reproducible decoding, and an OpenAI-compatible API server for local chat and coding agents. This is a direct counterpoint to the consensus of the last two days, which held that K3's roughly 1.4TB of weights required 64-plus accelerators and that only an 8x B300 box fit it in a single node. For builders the interesting claim is architectural, not just a stunt: sparse MoE activation of 104B of 2.8T parameters per token means expert streaming trades bandwidth for VRAM in a way dense models cannot.
↳ Follow the thread