Apple Is Talking to the Startup That Squeezes Giant AI Into an iPhone

Key points

• PrismML, a Caltech spinout backed by Khosla Ventures, publicly released a compressed AI model on Tuesday that fits on an iPhone 15 or newer.
• The startup says it reduced Alibaba's open-source Qwen model from roughly 54 GB to under 4 GB, a more than 13-times reduction in size.
• Apple is in early talks with PrismML to evaluate the technology for use on its devices, PrismML CEO Babak Hassibi told CNBC.
• PrismML raised a $16.25 million seed round in March 2025.
• The compressed models use 10 to 15 times less memory and generate responses 6 to 8 times faster than standard versions, according to the company.

A Silicon Valley startup says it has worked out how to squeeze a powerful AI model, the kind of software that drives chatbots like ChatGPT, down to a size small enough to run entirely on an iPhone. And Apple is reportedly interested.

PrismML, a company spun out of the California Institute of Technology, publicly released compressed versions of Alibaba's open-source Qwen model on Tuesday. Qwen normally takes up about 54 gigabytes of storage. PrismML says it cut that to under 4 gigabytes, while keeping all 27 billion of its internal data points, called parameters, intact.

CEO Babak Hassibi told CNBC that Apple and other unnamed companies are actively testing the startup's models, measuring speed, energy use and overall performance on real devices. He called the Apple discussions very early but said "things are progressing nicely."

The release landed one day after Apple opened its iOS 27 public beta, which gives iPhone owners their first broad look at the company's long-delayed overhaul of Siri.

Why does it matter whether AI runs on your phone?

Running AI directly on a device, rather than sending your request to a distant server and waiting for the answer to come back, is faster and keeps your data on your phone. That matters most when the information is personal: health records, medication schedules, photos. It also means certain features work without any internet connection at all.

Apple already handles some tasks locally, including translation and basic summarisation. More demanding requests go to the cloud. The company wants to flip that balance.

PrismML's trick is to drastically simplify how values are stored inside a model. A standard model stores each value using 16 possible states, like a very precise measuring scale. PrismML compresses each value down to just one or three possible states. Think of it as switching from a finely graded ruler to a simple on-off switch. That cuts memory use sharply.

The company says the trade-off is a small drop in accuracy. Factual recall weakens first. Skills like reasoning and maths hold up better.

Analysts welcomed the concept but urged caution. Tarun Pathak of Counterpoint Research said performance across millions of real-world queries, on thousands of different device combinations, is the true test. Phil Solis of IDC flagged battery life as the biggest open question: a model used continuously in the background could drain a phone even if it needs less memory.

For everyday iPhone users, nothing changes yet. This is a business negotiation, not a software update. But if Apple and PrismML reach an agreement, future Siri improvements could arrive faster, work offline and handle more sensitive information without it ever leaving your pocket.

PrismML is releasing two compressed model versions for free, compatible with iPhones, MacBooks and Nvidia-powered Windows PCs. Google's Gemma model is next in the company's pipeline.