A developer manages to bypass limits on Apple’s M4 chip for AI training and reopens the debate about local power in laptops
A new experiment of reverse engineering Apple’s M4 chip has put the focus back on the platform’s real capacity for artificial intelligence tasks beyond what Apple officially allows. According to the information released by the researcher who signs as @0x0SojalSec, it would have been possible to circumvent the M4’s software restrictions to use its power in model training, something that Apple does not openly enable on its equipment, where the Neural Engine is mainly oriented towards inference.
The publication maintains that this work has managed to take advantage of up to 15.8 TFLOPS of computing power in AI tasks within the M4, keeping all the flow in RAM to avoid writing to NAND and thus reduce latency. The most striking detail is that, according to its author, the achievement was not supported by common tools from the Apple ecosystem such as CoreML or Metal, nor did it resort to the GPU in the way a conventional application would, but rather to its own environment described as a MIL, a Model Intermediate Language created from scratch.
The statement must be read with caution. For now we are talking about a technical project disseminated by its authornot a feature validated or documented by Apple. Even so, the case is interesting because it points to something that has been hovering over the market for some time: the distance between the physical power of certain modern chips and the framework of use that their manufacturers allow at the software level. In the case of Apple, this especially affects devices with the M4 chip, a platform that was already born with a strong orientation towards Apple Intelligence and local AI loads.
The M4 would have more margin than what Apple officially allows it to use
According to the shared explanation, one of the keys to the system is how the software manages blocking or restarting the training process. The author mentions the use of the command exec() to restart execution and continue learning when the process needs to refresh state. It also highlights that by keeping data in RAM instead of constantly writing to storage, behavior is much smoother and faster.
What is relevant here is that Apple has designed its recent chips with great heterogeneous processing capacity, but reserves an important part of that potential for specific uses defined by its own software stack. If this experiment truly reflects what can be done outside of those official routes, the implicit message is that there is more local capacity available which Apple allows to exploit as standard on Mac or iPad.
That’s especially interesting at a time when the entire industry is trying to bring AI closer to the device. It is not just about running a pre-trained model, but about exploring the extent to which certain consumer equipment can also take on more ambitious loads, including training or tuning. And there comes an inevitable comparison with what we are seeing in other market players.
The battle for local AI is no longer limited to Apple
These types of experiments come just as manufacturers are pushing the local hardware argument for AI more strongly. At Computex 2026 we saw how NVIDIA presented the RTX Spark for laptops, processors with Arm CPU, unified memory and an integrated Blackwell GPU designed precisely to make AI a central element of the PC. There the discourse was very clear: bring agents, models and accelerated loads to the local team without always depending on remote infrastructure.
In this context, the case of the M4 reinforces a broader idea. The competition is no longer just about who has the best assistant or the best software integration, but also about how much local power can actually be activated and under what conditions. Apple continues to opt for a more controlled and closed approach, while other players try to sell platforms prepared from the beginning for more open or more intensive AI loads.
