Apple eyes brilliant math breakthrough to run giant AI models entirely on your iPhone
Apple might be partnering up with Caltech spinoff PrismML to run properly large language models on the iPhone.
PrismML is redefining the "large" in large language models. | Image by Apple
Apple is reportedly considering using a much larger LLM on the iPhone, recent reports claim. Cupertino has held talks with PrismML, Such a cooperation could allow Apple to put much more capable AI language models on the iPhone without it occupying as much storage as one normally does.
Optimizing LLMs could be the key to enhanced Apple Intelligence
Optimizing LLMs could be the key to enhanced Apple Intelligence
What does PrismML actually do? The company is known for building ultra-dense large-language models that deliver the same performance at a fraction of the size. That's essentially compressing the model and slimming down its hardware requirements so that it can run on much weaker hardware without sacrificing performance.
The company has reportedly been able to reduce the size of Alibaba's powerful Qwen 3.6 27B model to run on the iPhone 17 Pro. With 27 billion parameters, this model is more capable than the AFM 3 Core Advanced model that is found on the iPhone 17 Pro/Pro Max and has merely 20 billion parameters.
Much smarter Apple Intelligence in the future?
The key moment here is that Apple relies on a sparse architecture, which only allows between one and four billion parameters of the model to be active during data processing, whereas PrismML's solution allows all 27 billion parameters of Qwen 3.6B to be active at the same time, making this the largest ever AI model running natively on an iPhone, so far.
This means PrismML's solution is between 7 to 27 times better-performing on the same iPhone 17 Pro hardware than Apple's current solution.
This essentially means that Apple could potentially develop and deliver much more powerful Apple Intelligence features in the near future, ones that would prove to be extremely competitive, especially on the mobile landscape.
Is the AI bubble popping in the next couple of years?
How does PrismML achieve that? Why, through the wondrous magic of algebra, of course!
Normally, running an LLM with 27 billion params would require at least 64 GB of RAM (something no phone has), but PrismML relies on linear algebra to greatly reduce that. Instead of using complex fraction data, PrismML's framework relies on simplified 1-bit ternary weights, which are much more efficient than standard LLMs.
That's obviously a great breakthrough.
Does this mean that future iPhones will utilize Chinese LLM power?
Probably not. While Alibaba's large models are very competitive, I can only imagine the backlash Apple will get for relying on a Chinese AI model for future iterations of its Siri AI. PrismML's solution is likely merely a proof of concept.
If AI is here to stay, at least make it optimized
I'm wary of the AI as a whole, just as many other people surely are. Yet, the technology is certainly here to stay despite the not-so-welcome public adoption.
If that should be the case for our near future, then at least make it as efficient and optimized as possible. A modern iPhone already harbors more power than most people need, we might as well optimize as much of AI as possible.
Everything to offset what those large and noisy AI data centers are doing.
Popular stories
![T-Mobile is down for many users across the nation right now [UPDATED]](https://m-cdn.phonearena.com/images/article/181747-wide-two_350/T-Mobile-is-down-for-many-users-across-the-nation-right-now-UPDATED.webp)
Latest News
Things that are NOT allowed:
To help keep our community safe and free from spam, we apply temporary limits to newly created accounts: