Apple is looking for ways to shrink heavy artificial intelligence models so they can run directly on your phone. To do this, the company has been holding talks with PrismML, a Caltech spinout that specializes in squeezing massive neural networks into tiny packages. According to a report by The Information, Apple wants to run advanced intelligence tasks on its devices without paying massive cloud server bills
.
The startup recently made headlines by compressing Qwen 3.6, a 27 billion parameter model developed by Alibaba. Normally, a model of that scale requires about 54GB of space, making it impossible to fit on a smartphone. PrismML managed to shrink the file size to under 4GB. That is small enough to run on an iPhone 15 Pro or iPhone 17 Pro. The newly released model, called Bonsai 27B, runs with all of its parameters active at the same time.
Keeping AI local is a massive priority for Apple. Right now, complex Siri requests must go to remote servers. This introduces lag and raises privacy concerns. By running larger models directly on the iPhone, Apple can process complex text, reasoning, and coding tasks without an active internet connection. It also shields user data from outside servers. During an interview with CNBC, PrismML CEO Babak Hassibi confirmed that Apple has been testing their technology for speed and energy efficiency. He pointed out that the talks are still in a very early stage. Apple has declined to comment on the discussions.
The mathematics behind this compression rely on 1 bit and ternary architecture. Instead of storing complex data paths, the system simplifies weight values down to basic states. PrismML claims this can reduce memory requirements by 10 to 15 times while speeding up processing by 6 to 8 times. Power consumption also drops by 3 to 6 times.
There is a catch. Compressing a model so drastically does degrade its accuracy. The startup admits that factual memory takes a hit, though reasoning and mathematical logic hold steady.
Tech companies are eager to see if this technology can survive real world testing. Running heavy software on a phone can quickly drain the battery and cause overheating. If PrismML can prove its architecture is stable, it could change how mobile chips and data centers are designed. For now, the open source model is available for developers to test. It represents a major step toward a future where almost all of your daily AI needs are handled directly in your pocket.
