What it is
A proof-of-concept that runs transformer forward and backward passes directly on the Apple Neural Engine by calling reverse-engineered private APIs and generating MIL compute graphs, bypassing CoreML. It trains small models such as Stories110M and Qwen3-0.6B, keeps weight gradients and Adam updates on the CPU, and includes benchmarks for throughput, SRAM behavior, and INT8 quantization. Researchers probing NPU hardware on Apple Silicon can read it as a reference or fork it, but it is not a maintained framework.