Laya on Axera AX650 and AX8850 edge NPUs, under 70 ms

Validated Core ML ports of all Laya checkpoints. One short multilingual decision in 4.98 ms P50 on M3 Max with ANE FP16, at 2.78x better energy per decision than MLX FP16.

Swift package that reads a form in a running Mac app through the Accessibility API, asks an on-device model what belongs in each field, and types it in. ~1 ms per decision.

FluidInference ported Laya to CoreML with 99.5% of ops running on the ANE and benchmarked it at 3.7 ms per decision on an M5 Pro.

Ggml-based C++ runtime for all three checkpoints. Native tokenization, decision heads and JSON output; optimized FP16/BF16 CUDA on RTX with FP32 accumulation, plus Vulkan.