Qwen3.5 35B hits iPhone, 5.6 tok/sec

This is a strong proof-of-concept for phone-side MoE inference: the headline number is less important than the memory trick that makes it possible. If this holds up beyond a demo, local AI on mobile is moving from “tiny toy models” toward genuinely useful mid-sized models.

• –256-expert sparsity plus 4-bit quantization is exactly the kind of setup that can make a large MoE model mobile-feasible.
• –SSD-to-GPU streaming is the real engineering win here; it shifts the bottleneck from fit-to-RAM to bandwidth and scheduling.
• –5.6 tok/sec is not desktop-fast, but it is fast enough to feel interactive on a phone.
• –The 379B follow-up will be the more interesting stress test, because it will show whether the approach scales or just flatters the smaller model.
• –This is more infrastructure than model hype: the model is familiar, but the deployment path is the news.