Flash-iOS Ports Sparse MoE Inference to iPhone

Flash-iOS’s real breakthrough is architectural: treating the iPhone’s SSD as a slow expert reservoir instead of requiring the entire model in memory.

• –MoE sparsity lets the runtime load only selected experts for each token, dramatically reducing active memory requirements.
• –Parallel, page-aligned reads attack the main bottleneck, but fragmented SSD access still makes generation far slower than desktop inference.
• –iOS-specific entitlements, pread-only loading, and cleanup fixes are essential engineering work rather than superficial platform porting.
• –The transfer script and model-management UI make experimentation accessible to developers with physical devices.
• –This is compelling edge-inference research and open-source infrastructure, but not yet a comfortable daily-driver local LLM experience.