Flash-MoE: Running a 397B Parameter Model on a Laptop

A custom C/Metal inference engine that streams Qwen3.5-397B-A17B (397B parameter MoE) entirely from SSD on a MacBook Pro with 48GB RAM, achieving 4.4+ tok/s with full tool calling at 4-bit quantization — no Python, no ML frameworks. The key technique is on-demand expert weight streaming via parallel pread() from NVMe, sidestepping the RAM bottleneck of dense models. Demonstrates that production-quality local inference of frontier-scale models is achievable on consumer Mac hardware with careful systems engineering.