Developer Unlocks Hidden AI Training Power Inside Apple’s M4 Neural Engine
Apple’s M4 chip is already known for its impressive AI performance, but a new developer breakthrough suggests the silicon may be capable of far more than Apple officially allows. A researcher using the name 0x0SojalSec has reportedly reverse-engineered the Apple M4 Neural Engine and unlocked up to 15.8 TFLOPS of AI training performance that is normally inaccessible through Apple’s standard software tools.
The Apple Neural Engine is designed to accelerate artificial intelligence tasks across Macs and iPads, but its official use is largely limited to inference. In simple terms, inference means running AI models that have already been trained. Training new models, especially advanced machine learning systems such as transformers, is a much heavier process and is not something Apple typically exposes directly to developers through its consumer-facing software stack.
That is what makes this discovery so interesting. Instead of relying on Apple’s official CoreML or Metal frameworks, the developer built a custom Model Intermediate Language from scratch. This allowed direct interaction with the M4 Neural Engine in a way that bypassed Apple’s usual restrictions. According to the shared technical details, this custom approach made it possible to run full backpropagation and transformer training directly on the Neural Engine itself.
Backpropagation is one of the key processes behind AI model training. It allows a neural network to learn from errors and improve over time. Being able to run this process on Apple’s dedicated AI hardware is a major step beyond simply using the Neural Engine to execute pre-trained models.
The project also required several clever workarounds to keep the system running reliably. Since the M4 Neural Engine is not officially opened for this kind of workload, stability became a major challenge. If a training process stalled, the custom language could trigger a specific execution command that effectively respawned the process. This refreshed the current state and allowed the machine learning task to continue without bringing down the entire program.
Performance optimization was another key part of the breakthrough. To avoid slowdowns, the developer configured the process to run entirely through system RAM instead of relying on slower NAND flash storage. This helped keep the AI training workload moving quickly and reduced bottlenecks during intensive operations.
For owners of M4-powered Macs and iPads, the discovery highlights just how powerful Apple’s latest silicon really is. The M4 Neural Engine appears to have enough raw capability to handle demanding AI training tasks, even though Apple currently restricts that functionality through software.
This does not mean everyday users can suddenly train large AI models on their M4 devices with a simple app. The method is highly technical, experimental, and built outside Apple’s official development environment. However, it does show that the hardware has untapped potential that could become extremely valuable as on-device artificial intelligence continues to grow.
Apple has been pushing deeper into AI across its ecosystem, and the M4 chip plays a central role in that strategy. If the company ever chooses to officially open more Neural Engine capabilities to developers, future Macs and iPads could become far more useful for local AI development, private model training, and advanced machine learning workflows.
For now, this reverse-engineering achievement serves as a fascinating look beneath the surface of Apple’s tightly controlled hardware. The M4 Neural Engine is not just a tool for running AI features in the background. With the right low-level access and custom software, it appears capable of handling serious AI training workloads that Apple has not yet made officially available.
