Apple in talks with startup that shrinks AI models to run on an iPhone
Apple The startup’s CEO told CNBC that it is in talks with a small Silicon Valley company that says it can make powerful AI models small enough to run directly on the iPhone.
PrismML, a Khosla Ventures-backed spinoff of the California Institute of Technology, publicly released compressed versions of Alibaba’s open-source Qwen model on Tuesday. The company said it shrinks the model from roughly 54 GB to under 4 GB, allowing all 27 billion parameters to run on an iPhone 15 or newer model.
PrismML CEO Babak Hassibi told CNBC that Apple and other companies are evaluating the startup’s models and measuring speed, energy efficiency and performance on devices.
“They’re really evaluating our technology right now,” Hassibi said of Apple.
He characterized the discussions as very early and said it remains unclear where they will lead, but that “things are progressing nicely.”
Apple did not immediately respond to a request for comment.
Information previously reported on the PrismML breakthrough.
The release comes a day after Apple opened a public beta of iOS 27, giving iPhone owners their first broad access to the company’s long-delayed Siri overhaul. Apple is trying to make Siri more competitive with assistants OpenAI And anthropic while keeping more personal information and AI processing on the device.
The company’s approach could address one of the key limitations facing Apple’s AI strategy. The most capable models usually require a lot of memory and processing power to run on a smartphone.
Apple can send complex requests to cloud-based models, but running more AI directly on the iPhone would reduce the latency associated with sending data to a remote server, lower cloud computing costs and bolster the company’s privacy pitch. It also allows some features to work without an internet connection.
Smaller models could allow Apple to bring more demanding features to the iPhone, including computational photography, video creation and health or fitness tools that rely on sensitive personal data, said Carolina Milanesi, president and principal analyst at Creative Strategies.
“The more you can do on the device, the better,” he said, pointing to health and pharmaceutical data that users would like to keep private.
PrismML said it shrinks AI models by greatly simplifying the way internal information is stored, reducing each value from 16 bits to just one or three possible values. This significantly reduces the memory required to store and run the model.
Hassibi compared this to the chip industry’s shift from eight-bit to four-bit computing, but took it a step further.
The startup said the compressed models use 10 to 15 times less memory, produce six to eight times faster response, and consume three to six times less energy than traditional versions running on existing hardware.
However, Hassibi acknowledged that a compromise was possible. He said PrismML models often lose a few percentage points in overall performance, with actual recall weakening before skills like reasoning, math and coding.
PrismML releases two compressed versions of the model for free. They’re designed to run on everyday devices, including iPhones, MacBooks, and Nvidia-powered PCs.
The technology emerged from Hassibi’s research group at Caltech. The University owns the underlying patents and licenses them exclusively to PrismML. In March, the company raised a $16.25 million seed round with backing from Khosla Ventures and other investors.
Hassibi said GoogleThe open source Gemma model is next, followed by much larger models, including models from frontier labs that today often require data center hardware.
The technology could ultimately expand well beyond phones and laptops to robotics, autonomous systems and other products that need to make decisions quickly without relying on cloud connectivity, according to PrismML.
“It is very important that the intelligence is local and can work quickly,” he said.
Apple’s on-device advantage
Apple already runs parts of its AI system natively, including translation, some summaries, and features closely tied to personal information. More complex requests are routed to Apple’s private cloud infrastructure or external models.
Asymco founder Horace Dediu said Apple is likely trying to keep the vast majority of common Siri interactions on the device while reserving the most demanding tasks for the cloud.
He said that the advantage is not just using less memory, but also including a more capable model within the same physical limits.
“They’re trying to figure out how big and how smart of a model they can fit into the device,” said Dediu. Keeping common requests local gives Apple lower latency, greater privacy, and potentially lower licensing and cloud costs.
Apple may have an advantage in running these models because it co-designs the iPhone’s chips and software, giving it tighter control over how the AI works on the device.
But analysts warned that PrismML’s claims still need to be proven outside of controlled demonstrations.
Tarun Pathak, research director at Counterpoint Research, said that the model’s performance on long requests, battery consumption during multitasking and reliability across millions of requests will be critical.
“The ultimate test will be millions of queries, thousands of device combinations, and robust testing at scale,” Pathak said.
Phil Solis, who leads IDC’s research on client processors, said power consumption may be the biggest unanswered question. A model that is capable enough to be used frequently or constantly in the background for agent-like tasks can drain the phone’s battery even if it requires less memory.
What does it mean for chip demand?
PrismML’s launch also comes amid an intense debate about whether improvements in AI efficiency could ultimately reduce demand for memory chips and expensive data center infrastructure.
Memory has become one of the biggest constraints and costs in consumer electronics and AI servers. Morgan Stanley It estimates that Apple’s average cost of dynamic random access memory per bit could rise roughly 190% year over year in fiscal 2027, while NAND costs could rise nearly 180%. NAND is often used in flash drives and solid state drives.
The firm expects Apple to raise the starting price of similar iPhone 18 models by around $200 to maintain margins.
PrismML said its approach could allow a cloud model that normally requires eight GPUs to run on a single device, while also allowing models that once required servers to be ported to phones and laptops.
This can reduce the amount of memory or computing capacity required for a particular AI task. However, this does not mean that overall chip demand will necessarily decline.
Gil Luria, an analyst at DA Davidson, said that miniaturizing models will not eliminate the need for processors or memory. It could move more of these chips from data centers to phones and other devices.
“It doesn’t mean you won’t need a chip,” Luria said. “You’ll still need the GPU, and you’ll still need memory.”
He added that running AI on individual devices may actually be less efficient than using shared data center infrastructure because the chips in phones can sit idle most of the time.
Breakthroughs in efficiency could also lead to higher usage rather than lower spending, as cheaper, faster AI enables new products and encourages consumers to run models more frequently.
Still, the market has been quick to punish anything that suggests AI might need less memory than expected. Micron Shares fall after Google releases March report TurboQuant paper Although the stock will recover later, the model is about cutting memory usage without hurting performance.
PrismML’s public release gives casual users and investors a chance to test whether the claimed gains hold true outside the lab. For Apple, running more capable AI directly on the iPhone could help the company improve Siri without giving up the privacy and hardware integration that distinguishes its products.
“The combination of cloud and on-device AI can deliver a more complete, efficient and privacy-focused AI experience,” Counterpoint’s Pathak said. “Complex tasks will be offloaded to the cloud, while sensitive, latency-critical and privacy-related tasks will be executed on-device.”
WRISTWATCH: Apple sues OpenAI for alleged trade secret theft: Here’s what you need to know
