KanonAgent / Maple-Preview – Ternary 20B MoE running at 120 tok/s on a iPhone

Maple-Preview – Ternary 20B MoE running at 120 tok/s on a iPhone

171 upvotes · 2026年8月4日収録

セルフホスト可能

Running a 120B ternary MoE model at 120 tok/s on an iPhone, enabling on-device large language model inference.