Apple M4 Neural Engine reportedly unlocked for AI training without CoreML, Metal, or GPU support
Apple’s M4 chip is already known for its strong AI performance, but a new claim suggests that the silicon may be capable of far more than Apple officially allows developers to access. According to a developer and security researcher posting under the handle @0x0SojalSec, the Neural Engine inside the M4 has reportedly been used for full AI model training, not just inference.
That is a notable claim because Apple’s Neural Engine is typically exposed to developers for running pre-trained machine learning models. In other words, it is designed to help apps perform AI tasks quickly and efficiently, such as image recognition, language processing, and on-device intelligence features. Training new models, however, is a very different workload and is usually handled by GPUs or dedicated AI accelerators in more open computing environments.
The researcher says they managed to bypass the usual software limitations and communicate with the M4’s Neural Engine without relying on Apple’s standard developer tools. That means no CoreML, no Metal, and no GPU acceleration. Instead, the work was reportedly done using a custom Model Intermediate Language, or MIL, built from scratch.
Because Apple’s hardware stack is tightly controlled, the project required some unusual engineering. The researcher claims the custom system can recover when the training process stalls by using an “exec()” command to restart or respawn the process and continue training. In simple terms, this technique appears to help the training workflow refresh itself and keep running instead of crashing when it hits a blocked state.
Another key part of the claim is that the entire process runs in RAM rather than writing data to NAND flash storage. This matters because RAM is much faster than internal storage, and avoiding frequent writes can significantly improve performance while also reducing storage wear. Keeping the workload in memory reportedly helped the system run smoothly and at high speed.
The most eye-catching figure mentioned is 15.8 TFLOPS of raw AI performance from the M4’s Neural Engine. If accurate, that level of performance could make Apple’s latest chips far more useful for local AI development than previously assumed. It could also open the door for training smaller AI models directly on consumer hardware like a Mac or iPad, without needing expensive desktop workstations or high-end dedicated graphics cards.
This is especially interesting at a time when AI workloads are becoming more important across personal computing. Many users want to run and fine-tune AI models locally for privacy, speed, and cost reasons. If Apple Silicon can be pushed beyond inference and into training workloads, it could change how developers view Macs and iPads as AI development machines.
However, it is important to treat the claim cautiously for now. Apple has not officially supported this kind of Neural Engine access, and there is no confirmation that the method is stable, safe, or widely usable across different devices. Reverse-engineered approaches can also break with software updates or hardware changes, especially when they depend on private behavior that was never intended for public development.
The report also raises questions about future Apple chips. If this kind of access is possible on the M4, many will naturally wonder what could be achieved on newer Apple Silicon, including the M5 generation. Still, it remains unclear whether the same custom MIL approach would work on newer chips or whether Apple has changed the underlying behavior of its Neural Engine.
For now, the claim highlights a fascinating possibility: Apple may have built far more AI capability into its chips than developers can currently use through official channels. Whether this becomes a practical path for local AI training or remains an impressive experimental breakthrough will depend on further testing, verification, and whether the broader developer community can reproduce the results.
