apple The company’s CEO told CNBC that the company is in talks with a small Silicon Valley company that says it can scale down its powerful artificial intelligence models to the point where they can run directly on the iPhone.
PrismML, a Caltech spinout backed by Khosla Ventures, released a compressed version of Alibaba’s open source Qwen model to the public on Tuesday. The company said it has reduced the model’s capacity from about 54 GB to less than 4 GB, allowing all 27 billion parameters to run on iPhone 15 and later.
Babak Hassibi, CEO of PrismML, told CNBC that Apple and other companies are evaluating the company’s models to measure the speed, energy efficiency and performance of their devices.
“They’re seriously evaluating our technology right now,” Hashibi said of Apple.
He characterized the talks as being at a very early stage and said that while it remained unclear what direction the talks would take, “things are progressing well.”
Apple did not immediately respond to a request for comment.
The Information previously reported on PrismML’s breakthroughs.
The release comes a day after Apple rolled out the public beta of iOS 27, giving iPhone owners broad access to the company’s long-delayed overhaul of Siri for the first time. Apple is trying to make Siri even more competitive with its assistant OpenAI and human While keeping more personal information and AI processing on your device.
The company’s approach could address one of the core constraints facing Apple’s AI strategy. The most powerful models typically require too much memory and processing power to run on a smartphone.
While Apple can send complex requests to cloud-based models, running more of the AI directly on the iPhone would reduce the delays associated with sending data to remote servers, lowering the cost of cloud computing and supporting the company’s privacy pitch. It also allows certain features to work without an Internet connection.
Carolina Milanesi, president and principal analyst at Creative Strategies, said a smaller model could allow Apple to include more demanding features on the iPhone, such as computational photography, video generation, and health and fitness tools that rely on sensitive personal data.
“The more you can do on the device, the better it will be,” she said, pointing to health and medication data that users want to keep private.
PrismML said it significantly simplifies how internal information is stored, shrinking AI models by reducing each value from 16 bits to one or three possible values. This significantly reduces the memory required to store and manipulate models.
Hassibi compared this to the chip industry’s transition from 8-bit to 4-bit computing, but takes it a step further.
According to the company, this compression model uses 10 to 15 times less memory, generates responses six to eight times faster, and consumes three to six times less energy than traditional versions that run on existing hardware.
However, Hasibi acknowledged there are trade-offs. He said PrismML’s models typically suffer from a several percent drop in overall performance, with reduced ability to reproduce facts before skills such as reasoning, math, and coding.
PrismML releases two compressed versions of the model for free. They are designed to run on everyday devices like iPhones, MacBooks, and Nvidia-powered PCs.
The technology comes from Hashibi’s research group at the California Institute of Technology. The university owns the underlying patents and has exclusively licensed them to PrismML. In March, the company raised $16.25 million in a seed round backed by Khosla Ventures and other investors.
Hashibi said google‘s open source Gemma model is next in the pipeline, followed by larger models, such as Frontier Labs’ models, which today typically require data center hardware.
According to PrismML, the technology could eventually extend far beyond phones and laptops to robots, autonomous systems, and other products that require rapid decision-making without relying on cloud connectivity.
“It’s very important that the intelligence is local and can be executed quickly,” he said.
Apple’s on-device benefits
Apple already runs some of its AI systems locally, including translations, some summaries, and functions closely related 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 looking to keep the majority of common Siri interactions on-device, while reserving the most demanding tasks in the cloud.
The advantage, he says, is not just that it uses less memory, but that it allows you to fit a higher-performance model within the same physical limitations.
“They’re trying to figure out how big a model and how smart a model they can fit into a device,” Dediu said. By keeping common requests local, Apple could potentially enjoy lower latency, better privacy, and lower licensing and cloud costs.
Apple could have an advantage in running these models because it co-designed the iPhone’s chips and software and has tighter control over how the AI runs on the device.
But analysts cautioned that PrismML’s claims would need to be proven outside of a controlled demonstration.
Tarun Pathak, research director at Counterpoint Research, said the model’s performance on long prompts, battery consumption when multitasking, and reliability over millions of requests will be important.
“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 client processor research, said power consumption may be the biggest unanswered question. Models that have enough functionality to be used frequently, or that can be used continuously in the background for agent-like tasks, may drain your phone’s battery even though they require less memory.
What it means for chip demand
The release of PrismML comes amid intense debate over whether improved AI efficiency will ultimately reduce demand for memory chips and expensive data center infrastructure.
Memory is one of the biggest constraints and costs across consumer electronics and AI servers. morgan stanley Apple estimates that the average cost per bit of dynamic random access memory will increase by approximately 190% year-over-year in 2027, and the cost of NAND will increase by approximately 180%. NAND is commonly used in flash drives and solid state drives.
The company expects Apple to raise the starting price of comparable iPhone 18 models by about $200 to ensure profits.
PrismML said its approach allows cloud models that typically require eight GPUs to run on one GPU, while potentially allowing models that previously required servers to be moved to mobile phones and laptops.
This can potentially reduce the amount of memory or computing power required for certain AI tasks. But that doesn’t necessarily mean overall demand for chips will decline.
Gil Luria, an analyst at DA Davidson, said shrinking the model doesn’t eliminate the need for processors or memory. They could simply move more of these chips out of data centers and into phones and other devices.
“It’s not that you don’t need a tip,” Luria said. “We’re still going to need GPUs and memory.”
He added that running AI on individual devices may actually be less efficient than using shared data center infrastructure, as the chips in mobile phones can sit idle much of the time.
Efficiency breakthroughs could also lead to increased usage rather than reduced spending, as cheaper and faster AI enables new products and encourages consumers to run models more frequently.
Still, the market has been quick to punish any suggestion that AI might require less memory than expected. micron The stock price plummeted in March after Google released its TurboQuant paper on reducing memory usage without compromising model performance, but the stock has since recovered.
The public release of PrismML gives the public and investors the opportunity to test whether its claimed benefits hold up outside the lab. And for Apple, running better AI directly on the iPhone could help improve Siri without giving up the privacy and hardware integration that are hallmarks of its products.
“By combining cloud and on-device AI, we can deliver a more complete, efficient, and privacy-focused AI experience,” said Counterpoint’s Pathak. “Complex tasks are offloaded to the cloud, while sensitive, latency-sensitive, and privacy-related tasks are performed on-device.”
WATCH: Apple sues OpenAI for trade secret theft: Here’s what you need to know
