How PrismML claims to shrink AI models
PrismML, which grew out of research at the California Institute of Technology, unveiled compressed versions of Alibaba’s open-source Qwen model on Tuesday. The company said it cut the model’s size from around 54 GB to under 4 GB, enabling all 27 billion of its parameters to operate on an iPhone 15 or newer, according to CNBC.
Chief executive Babak Hassibi told CNBC that Apple and several other firms are currently testing the startup’s models for speed, energy use and overall performance. “They’re really evaluating our technology right now,” Hassibi said. He described the talks as very preliminary but added that “things are progressing nicely.”
Why on-device AI matters for Apple
The development lands a day after Apple opened public beta testing for iOS 27, which includes its long-awaited redesign of Siri.
Apple has been working to make the assistant more competitive with rivals from OpenAI and Anthropic, while keeping as much data and processing as possible on the device itself rather than in the cloud.
Running larger AI models locally could ease one of Apple’s biggest technical constraints, since the most capable systems typically demand more memory and processing power than a smartphone can normally provide. Doing so on-device would cut latency, reduce cloud costs and reinforce Apple’s privacy positioning, while also allowing some features to function offline.
According to CNBC report, PrismML said its method works by simplifying how a model’s internal values are stored, reducing each figure from 16 bits down to as few as one or three possible values. Hassibi compared the approach to the semiconductor industry’s shift from eight-bit to four-bit computing, adding that PrismML “takes it a step further.”
PrismML says its compressed models use up to 15 times less memory, run six to eight times faster and consume up to six times less energy, though Hassibi acknowledged a modest drop in performance, particularly in factual recall rather than reasoning or coding ability.
What it could mean for chip demand
The announcement arrives amid growing debate over whether such efficiency gains might dent demand for memory chips and datacentre hardware. Morgan Stanley has projected Apple’s memory costs could rise sharply in fiscal 2027, potentially pushing iPhone prices higher.
Hassibi said Google’s Gemma model is next for compression, followed eventually by larger frontier models that currently require datacentre-scale hardware. “It’s very important that the intelligence be local and that it can run fast,” he said.
