Someone is running K3's 2.8T parameters on a single M1 Max by streaming experts over HTTP.

deltafin, created July 28 and already at 243 stars, streams 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 server. This directly contradicts the consensus from two days ago that K3's ~1.4TB required 64+ accelerators. The architectural claim is the interesting part: sparse MoE activating 104B of 2.8T per token means expert streaming trades bandwidth for VRAM in a way dense models can't.