Apple (AAPL) Explores Game-Changing Tech to Squeeze Massive AI Into iPhones

Key Highlights

Table of Contents

Apple is reportedly discussing potential collaboration with Caltech-backed PrismML for on-device AI compression capabilities
PrismML successfully compressed Alibaba’s Qwen 3.6 model with 27 billion parameters from 54 GB to under 4 GB
The compression maintains all 27 billion parameters active concurrently on iPhone 17 Pro hardware
Apple’s existing premium Siri capabilities depend on cloud-based Google Gemini powered by Nvidia infrastructure
Khosla Ventures led PrismML’s $16.25M seed funding round launched this year

Apple (AAPL) shares climbed 0.44% on Wednesday following reports that the tech giant is engaged in discussions with PrismML, an artificial intelligence startup that originated from the California Institute of Technology.

Apple Inc., AAPL

According to initial reporting by The Information, the conversations focus on PrismML’s groundbreaking capability to execute a 27-billion-parameter artificial intelligence model entirely on an iPhone 17 Pro device — eliminating the need for cloud connectivity.

PrismML managed to compress Alibaba’s publicly available Qwen 3.6 model from approximately 54 gigabytes to below 4 gigabytes. This represents a size reduction exceeding 90%.

What distinguishes this compression technique from competing approaches is its performance retention. While most compressed models compromise accuracy to accommodate device constraints, PrismML maintains it doesn’t make such trade-offs.

The company accomplishes this through ultra-compact 1-bit and ternary weight structures, which slash memory demands by as much as 14 times and execute up to 8 times faster compared to conventional models.

Traditional on-device AI implementations only engage a portion of their parameters simultaneously — an approach known as sparse architecture. PrismML maintains all 27 billion parameters functioning together.

This enables the device-resident model to manage sophisticated operations including reasoning, autonomous agents, and software development, as stated by the startup.

The publicly accessible version of this model is scheduled for release next Tuesday.

Apple’s Reliance on Cloud Infrastructure

Currently, Apple’s most sophisticated Siri functionalities — unveiled at WWDC in June — depend on Google’s Gemini models operating on Nvidia processors within Google Cloud infrastructure.

Apple has deployed one new on-device model containing 20 billion parameters, though only 1 to 4 billion remain active simultaneously within its sparse configuration.

PrismML’s comprehensive-parameter methodology would represent a significant advancement. It would enable Apple to execute more demanding AI operations completely on-device, reducing cloud expenditures and maintaining user information away from third-party servers.

Background on PrismML

PrismML was established by Babak Hassibi, a professor specializing in electrical engineering at Caltech. The underlying mathematical research was conducted at the institution.

Caltech retains patent ownership for the fundamental technology while granting exclusive licensing rights to PrismML.

The company secured a $16.25 million seed investment earlier this year, with Khosla Ventures leading the round — the same investment firm that provided OpenAI’s initial venture capital funding.

Neither Apple nor PrismML have issued public statements regarding the potential acquisition discussions.

While competitors including Microsoft, Amazon, and Meta are investing hundreds of billions constructing data center capacity, Apple’s approach has persistently favored on-device AI processing.

A partnership with PrismML would advance that strategy beyond anything Apple has implemented previously.

✨ Limited Time Offer