Espresso runs transformers on Apple Neural Engine

Espresso compiles MIL programs directly to Apple's Neural Engine via reverse-engineered private APIs, bypassing CoreML entirely. On an M3 Max, it achieves 1.08 ms per token (926 tok/s) for a 6-layer transformer, 4.76x faster than CoreML and 11x faster than llama.cpp Metal. The Swift library supports M1 through M4 chips on macOS 15+.