ResearchBenchmarksInference Latency on Qwen 30B on M2 UltraFollow0.572Wall Time (s)NPUMoE0.549120.703560.8581.01244Apr 20, 2026Evaluation ResultsMethodMethodLinksWall Time (s)Expansion Time (s)Attention Time (s)Speedup (x)NPUMoEMode=oursMode=ours2026.040.5720.4640.0861coremlMode=coremlMode=coreml2026.041.1390.9780.0951.99anemllMode=anemllMode=anemll2026.041.1440.9880.092