llama-modchip Runs a 284B DeepSeek V4-Flash Model in 57GB of RAM on a CPU-Only 64GB Box, Hits 9-14 Tok/s

An Apache 2.0 project called llama-modchip that trended to 417 points on Hacker News fits DeepSeek V4-Flash's 284B weights into a 64GB CPU-only machine using a custom 1.71 bits-per-weight 'Fractal-64' quantization (cutting 82.5GB to 60.73GB), AVX-512 register tiling, and a Stockfish-PVS speculative drafter. Reported throughput: 44.4 tok/s prefill and 9.1-14.7 tok/s generation at 256K context on Zen 5, Apple Silicon and ARM64 - setting a new practical bar for GPU-less inference of frontier MoE models.