Apple in talks with PrismML to bring 27B-parameter AI to iPhone

Apple is in discussions with Caltech spinout PrismML to deploy a 27-billion-parameter AI model directly on the iPhone, a breakthrough that could eliminate the need for cloud processing in advanced AI tasks.

The startup successfully compressed Alibaba's open-source Qwen 3.6 model — which normally requires 54 gigabytes of memory — to less than 4 gigabytes, keeping all 27 billion parameters active simultaneously. Most AI models running on smartphones today activate only a few billion parameters at a time to manage power and heat constraints.

PrismML achieved the compression using ultra-dense 1-bit and ternary weight architectures, reducing memory footprint by as much as 14 times while running up to eight times faster than conventional compressed models, according to the company. The startup plans to release its open-source model on July 14.

The technology could reshape Apple's AI strategy. At WWDC, Apple introduced a revamped Siri architecture that still relies on sending data to Google's Gemini models in the cloud for complex tasks. On-device AI would offer total user privacy, faster response times, and zero dependence on cellular connectivity — while saving Apple the cost of cloud server infrastructure that rivals like Meta, Microsoft and Amazon are spending hundreds of billions to build.

Why on-device AI matters for investors

Apple has already held meetings with PrismML about potential applications, according to people familiar with the talks. The company's biggest secretive AI acquisition was Q.ai, a startup that reportedly went for $2 billion. Bringing PrismML's technology in-house would give Apple a differentiated capability that no major smartphone competitor currently offers.

The technical hurdle has been memory bandwidth. A 27-billion-parameter model at 16-bit precision requires about 54 GB of memory — far beyond what any smartphone can accommodate. PrismML's approach uses 1-bit and ternary weights that reduce each parameter to one or two bits instead of 16, slashing the memory requirement while preserving benchmark performance.

For context, Apple's A18 Pro chip in the iPhone 17 Pro has a unified memory architecture shared between CPU and GPU, with total system memory of 8 to 12 gigabytes depending on configuration. Running a full 27-billion-parameter model locally would have been impossible without the compression technique PrismML developed.

Competitive implications

The move would put pressure on Samsung and Google to accelerate their own on-device AI deployments. Samsung has been developing its Gauss models, while Google's Pixel lineup runs Tensor chips optimized for the company's Gemini Nano. Neither has demonstrated the ability to run a model of this scale entirely on-device with all parameters active.

Apple shares rose 0.88 percent to $313.39 on Wednesday. The company trades at roughly 30 times forward earnings. If PrismML's technology delivers on its claims, the resulting upgrade cycle for iPhones could be significant — on-device AI capabilities of this magnitude are widely considered the next catalyst for smartphone replacement demand.

This article is for informational purposes only and does not constitute investment advice.