Flash-MoE streams 35B models onto iPhone

Flash-MoE turns storage bandwidth into usable model capacity, making genuinely large local models viable on mobile hardware.

• –Qwen3.5-35B-A3B has 35B total parameters but only about 3B active per token, making it an ideal SSD-streaming target.
• –The trade-off is clear: lower RAM requirements come with per-token storage I/O, latency sensitivity, and thermal constraints.
• –Its pure C, Objective-C, and Metal implementation keeps the runtime lightweight and tightly aligned with Apple Silicon.
• –iOS ports require specialized memory entitlements, pread-based loading, and model-management work, so this is still an engineering demo rather than a turnkey consumer app.
• –The approach could make private, offline inference far more capable as mobile flash storage and bandwidth improve.