A developer has found a way to bypass Apple's restrictions on the M4 chip's Neural Engine, allowing on-device AI model training on MacBooks.
Security researcher @0x0SojalSec announced on X that they reverse-engineered the limitations, unlocking up to 15.8 TFLOPS of compute power for full training workloads, including backpropagation on transformer models.
>>> Asus Dawn 7S Ryzen Edition Launches in China with Ryzen AI 5 330
Normally, the Neural Engine in M4 chips is used only for running pre-trained models efficiently. Apple markets it as delivering up to 38 TOPS for inference only.
The developer built a custom Model Intermediate Language (MIL) to communicate directly with the Neural Engine, bypassing Apple's Core ML or Metal frameworks.
>>> HyperOS 4 Might Debut on Xiaomi's Next Foldable, Not the 18 Series
The process keeps everything in RAM to avoid slow storage writes, making it fast and stable.
They also used the exec() command to restart the process when training gets stuck, allowing checkpointing and continuation.
The project is shared on GitHub with minimal extra dependencies.
>>> Stable Android 17 Rolling Out to Google Pixels: Eligible Devices, New Features
Early tests show transformer training steps completed in milliseconds on M4 chips.
This could turn regular Macs and iPads into capable local training machines for smaller models, reducing reliance on cloud services.
>>> Qualcomm unveils Snapdragon Reality Elite with 60% faster GPU, 160% better NPU
The discovery sparks discussions about Apple's tightly controlled hardware-software approach and reveals hidden potential in the chips.
