Why did Apple have to find a small startup to solve the AI ​​problem right on the iPhone?

There is news that I have read and reread several times over the past few hours: Apple, a company with enough money and enough people to build any AI model on its own, is sitting in a meeting with a new startup spun out of Caltech called PrismML, just to find out how they managed to cram an open source AI model from Alibaba, weighing up to 27 billion parameters, into the memory of an iPhone 17 Pro. This is notable not just because of the number, but because it reveals that Apple is recognizing a weakness that it has not yet resolved.

Who is PrismML, and what does it do?

PrismML is a company spun out of the California Institute of Technology (Caltech), funded by Khosla Ventures, and they have quietly emerged from anonymity since April 2026 with a model line called Bonsai. What makes PrismML different from most other AI companies is its extreme approach to model compression: instead of training a large model and then trying to quantize it (quantization: a technique to reduce the number of bits used to store each weight of the neural network), PrismML trains the model from the beginning in the form of ternary weights, meaning each weight can only take on one of three values: -1, 0, or +1.

PrismML, a startup spun out from Caltech, attracted attention when Apple discussed bringing its 27 billion parameter model to mobile devices. This is the point I find most worth mentioning technically. In the history of model compression, squeezing a neural network down to 1 or 2 bits per weight has almost always ruined performance, especially on complex inference tasks, because rounding numbers after training erases the subtle representations the model has learned. What PrismML calls Ternary Bonsai avoids this problem by not going the "post-compression" route but building a 1-bit architecture right from the mathematical foundation, incorporating group quantization techniques: every 128 weights share a common scale factor in FP16 format, so that each weight is essentially {-s, 0, +s} instead of just having three absolute fixed values.

An interesting comparison: discrete architecture and solid architecture

What I find more interesting than the compression story is the difference between PrismML's model and Apple's own current on-device model, called AFM 3 Core Advanced, which is running features like Siri AI's more expressive voice and system-wide dictation on the iPhone 17 Pro and iPhone Air. AFM 3 Core Advanced has 20 billion parameters, but uses a mixed architecture of experts, meaning that at a time only about 1 to 4 billion parameters are actually active, the rest lie dormant waiting to be called when needed. Meanwhile, the Qwen 3.6 model that PrismML compresses has a solid architecture, meaning all 27 billion parameters are active at the same time, with no dormant part.

To put it simply, in the Sparse model, only a few parameters are selected, while with Dense, all are active at the same time. In other words, Apple is going in the direction of "many parameters but only waking a small portion at a time" to save computation, while PrismML proves that with strong enough compression, a dense model with all the parameters active at the same time can still fit on mobile hardware, with the advantage that every parameter is utilized full time. These are two completely different design philosophies to solve the same problem: bringing powerful AI down to a chip in your pocket, and it seems that Apple wants to learn both directions instead of just going one way.

Why this is more important to Apple than just a meeting

I think the real motivation behind this meeting is not just technological curiosity. Currently, when an Apple Intelligence task exceeds on-device processing capacity, it will be pushed to Apple's Private Cloud Compute infrastructure, which means servers using Apple Silicon chips located in their own data centers. This is expensive, and although Apple always emphasizes the security element of Private Cloud Compute, in essence, handling everything right on the device is always a safer and cheaper option if done. If Apple can run 27 billion parameter models right on the latest A-series chip without sacrificing response speed, they will not only reduce server operating costs but also have a familiar marketing argument: privacy, this time tied to truly stronger AI features, not just slogans.

The story of operating on-device AI models is a painful problem that Apple has been trying to solve for the past few years.