Apple evaluates startup tech to run full AI models on iPhone

Apple explores breakthrough compression for on-device AI

Apple is evaluating technology from PrismML, a Caltech spinout, that could enable full-scale AI models to run directly on iPhones without cloud connectivity. The startup's CEO Babak Hassibi confirmed to CNBC that Apple and other companies are testing PrismML's models for speed, energy efficiency, and on-device performance.

PrismML publicly released compressed versions of Alibaba's open-source Qwen model this week, demonstrating it reduced the model from approximately 54GB to less than 4GB while preserving all 27 billion parameters. The compressed model can run on iPhone 15 and newer devices.

Hassibi characterized the discussions with Apple as early-stage, noting "things are progressing nicely" but the outcome remains uncertain. Apple did not respond to requests for comment.

How the compression works

PrismML's approach drastically simplifies how AI models store internal information, reducing each value from 16 bits to just one or three possible values. This technique cuts memory requirements by 10 to 15 times compared to conventional versions, according to the company.

The compressed models generate responses six to eight times faster and consume three to six times less energy than standard versions on existing hardware. However, Hassibi acknowledged a trade-off: the models typically lose a few percentage points of overall performance, with factual recall weakening before skills like reasoning and coding.

The technology emerged from Hassibi's research group at Caltech, which owns the underlying patents and licenses them exclusively to PrismML. The startup raised $16.25 million in seed funding in March from Khosla Ventures and other investors.

Why it matters

Running capable AI models directly on iPhones addresses a core constraint in Apple's AI strategy. On-device processing eliminates the latency of cloud requests, reduces computing costs, strengthens privacy claims, and enables features to work offline. This capability becomes critical as Apple attempts to make Siri competitive with assistants from OpenAI and Anthropic while maintaining its privacy positioning.

Carolina Milanesi of Creative Strategies noted that smaller models could enable Apple to move demanding features like computational photography, video generation, and health tools that process sensitive personal data entirely onto devices. "The more you can do on device, the better it is," she said.

Questions remain about real-world performance

Analysts cautioned that PrismML's claims require validation beyond controlled demonstrations. Tarun Pathak of Counterpoint Research emphasized that performance on lengthy prompts, battery consumption during multitasking, and reliability across millions of requests will prove critical.

Phil Solis of IDC highlighted power consumption as a major open question. A model capable enough for frequent use or continuous background operation could drain battery life even with reduced memory requirements.

Implications for chip demand

PrismML's release enters an ongoing debate about whether AI efficiency gains will reduce demand for memory chips and datacenter infrastructure. The startup claims its approach could allow a cloud model requiring eight GPUs to run on one, or move server-dependent models onto phones and laptops.

However, Gil Luria of D.A. Davidson argued that shrinking models won't eliminate chip demand but rather shift it from datacenters to consumer devices. He noted that running AI on individual devices can actually be less efficient than shared datacenter infrastructure because chips in phones often sit idle.

Morgan Stanley estimates Apple's average DRAM cost per bit could rise roughly 190% year over year in fiscal 2027, with the firm expecting Apple to raise iPhone 18 starting prices by approximately $200 to protect margins.

PrismML plans to release compressed versions of Google's Gemma model next, followed by much larger frontier models. The company is making two compressed versions of the Qwen model available for free, designed to run on iPhones, MacBooks, and Nvidia-powered PCs.

These details were first reported by CNBC.