On-device AI agents hit a hard memory limit. Apple's new architecture routes around it.

Apple’s AFM 3 models aim to bypass on-device DRAM limits by storing a 20B-parameter model’s weights in NAND flash and loading selected experts into DRAM once per prompt, activating about 1B to 4B parameters per task. Apple has not clarified offload rules or key performance metrics, with benchmarks expected in a summer report.