Apple in talks with Khosla Ventures-backed PrismML to shrink AI models for iPhone: Report
Apple is reportedly evaluating Khoslaventures-backed startup PrismML technology. The startup says it can make powerful AI models small enough to run directly on an iPhone and use up to 15 times less memory. CNBC It was reported on Tuesday, July 14.
How does PrismML claim to miniaturize 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 reduced the size of the model from approximately 54 GB to less than 4 GB, ensuring that all 27 billion parameters work on an iPhone 15 or newer model. CNBC.
Chief executive officer Babak Hassibi said: CNBC It is stated that Apple and many other companies are currently testing the startup’s models in terms of speed, energy usage and overall performance. “They are really evaluating our technology right now,” Hassibi said. He described the talks as very preliminary but added that “things are progressing nicely.”
Why is on-device artificial intelligence important for Apple?
The development comes a day after Apple launched public beta testing for iOS 27, which includes the long-awaited Siri redesign.
Apple is working to make the assistant more competitive with rivals in OpenAI and Anthropic, while also keeping as much data and processing as possible on the device itself rather than in the cloud.
Running larger AI models natively could alleviate one of Apple’s biggest technical constraints, as the most capable systems often require more memory and processing power than a smartphone can normally provide. Doing this on-device will reduce latency, lower cloud costs, and strengthen Apple’s privacy position, while also allowing some features to work offline.
Accordingly CNBC According to the report, PrismML said its method works by simplifying how a model’s internal values are stored, reducing each digit from 16 bits to one or three possible values. Hassibi compared the approach to the semiconductor industry’s transition from eight-bit to four-bit computation, adding that PrismML “takes that one 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; however, Hassibi acknowledged a modest decline in performance, particularly in actual recall rather than reasoning or encoding ability.
What could it mean for chip demand?
The announcement comes amid growing debate about whether such efficiency gains will reduce demand for memory chips and data center hardware. Morgan Stanley has predicted that Apple’s memory costs could rise sharply in fiscal 2027, potentially driving up iPhone prices.
Hassibi said Google’s Gemma model is the next model in terms of compression, followed by larger edge models that now require data center-scale hardware. “It is very important that the intelligence is local and can work quickly,” he said.
