For a long time, Apple's Neural Engine (ANE) has been tightly locked behind the "inference-only" curtain. But today, in 2026, that wall has been torn down. Recently, engineer Manjeet Singh, in collaboration with Claude AI, successfully reverse-engineered the computational secrets of the M4 chip's ANE, proving to the world: your Mac mini is not just for "raising lobsters," it can even be used directly to train Transformer models!

The core of this breakthrough lies in bypassing the bulky CoreML framework. With the assistance of Claude, Manjeet Singh delved into the mysteries of the MIL language and E5 binary, successfully achieving direct control over the ANE hardware. The experimental data is shocking: running a single-layer Transformer on the M4 chip achieves a peak energy efficiency of up to 6.6 TFLOPS/W, which is 80 times higher than the professional-grade A100 GPU and more than 50 times higher than the H100.

For a long time, the industry generally believed that NPU could not handle training tasks because of insufficient hardware. But this "brute-force cracking" revealed the truth: hardware has never been the bottleneck, it's Apple's software restrictions that are the real issue. Now, developers can already complete the full training of the Stories110M model on the Mac mini, with the entire system's power consumption as low as less than one watt.

This means that the threshold of the AI revolution is undergoing a dramatic change. The costly computing bills that once ran into tens of thousands of dollars now seem like a joke in front of the extreme energy efficiency of the M4 chip. For independent developers and home laboratories, expensive GPU clusters are no longer the only option. The small machine on your desk is turning into a supercomputer capable of low-cost iteration of large-scale models.

Although utilization still has room for improvement and there are significant engineering challenges, the door has now been opened. As developers put it, this kind of human-computer collaborative reverse exploration has shown us the dawn of edge-side AI training. In the future, the MacBook in your hands may no longer just be a consumer tool, but a private brain that evolves anytime, anywhere.