Apple is in talks with a startup that is scaling AI models to run on the iPhone

an apple is in talks with a small Silicon Valley company that says it can scale artificial intelligence models powerful enough to run directly on the iPhone, the company’s CEO told CNBC.

PrismML, a Khosla Ventures-backed spinout from the California Institute of Technology, released publicly compressed versions of Alibaba’s open Qwen model on Tuesday. The company said it has reduced the model from about 54 GB to less than 4 GB, allowing all 27 billion parameters to work on the iPhone 15 or newer.

PrismML CEO Babak Hassibi told CNBC that Apple and other companies have been testing launch models and measuring their speed, power efficiency and performance on devices.

“They’re really testing our technology right now,” Hassibi said of Apple.

He pointed out that the negotiations are still very early and it is not yet clear where they will lead, but “things are going well.”

Apple did not immediately respond to a request for comment.

Information previously reported the success of PrismML.

The release comes one day after Apple opened the public beta of iOS 27, giving iPhone owners their first broad access to the company’s long-awaited overhaul of Siri. Apple is trying to make Siri more competitive with assistants OpenAI again Anthropic while keeping more personal information and AI processing on the device.

The company’s approach could address one of the central issues facing Apple’s AI strategy. More capable models often require more memory and processing power to run on a smartphone.

Apple can send complex requests to cloud-based models, but using more AI directly on the iPhone can reduce the delay associated with sending data to a remote server, lower the cost of cloud computing and support the company’s privacy policy. It may also allow certain features to work without an internet connection.

Carolina Milanesi, president and chief analyst at Creative Strategies, said smaller models could allow Apple to move much-needed features to the iPhone, including computerized photography, video production and health or fitness tools that rely on sensitive personal data.

“The more you can do on a device, the better,” he said, pointing to health and medication data users might want to keep private.

PrismML is said to shrink AI models by greatly simplifying how their internal information is stored – reducing each value from 16 bits to one or three possible values. That greatly reduces the memory required to store and run the model.

Hassibi compared it to the chip industry’s move from 8-bit to four-bit computing, but takes it a step further.

The startup said the compressed models use between 10 and 15 times less memory, generate responses six to eight times faster and consume three to six times more power than conventional versions running on existing hardware.

Hassibi acknowledged that there are trade-offs, however. PrismML models typically lose a few percent of their overall performance, and factual recall is weak before skills like reasoning, math and coding, he said.

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 comes from Hassibi’s research group at Caltech. The University owns the underlying copyrights and licenses them exclusively to PrismML. In March, the company raised a $16.25 million seed round backed by Khosla Ventures and other investors.

Hassibi said GoogleThe open source Gemma model is next in the pipeline, followed by much larger models, including those from frontier labs that today often require datacenter hardware.

The technology, according to PrismML, could eventually extend beyond phones and laptops to robots, autonomous systems and other products that need to make quick decisions without relying on cloud connectivity.

“It is very important for intelligence to be local and able to run quickly,” he said.

Apple’s profit on the device

Apple already uses parts of its AI system in place, including translation, summarization and features closely linked to personal information. Complex applications are sent to Apple’s private cloud infrastructure or external models.

Horace Dediu, the founder of Asymco, said that it is possible that Apple is trying to keep most of the normal conversations of Siri on the device while keeping the most difficult tasks in the cloud.

The advantage is not just using less memory, he said, but fitting a more capable model within the same physical parameters.

“They’re trying to figure out how big a model and how smart a model they can fit into the device,” Dediu said. Keeping standard requests local gives Apple lower latency, greater privacy and potentially lower licensing and cloud costs.

Apple may benefit from using these models because it designs the iPhone’s chips and software together, giving it tighter control over how the AI ​​works on the device.

But analysts cautioned that PrismML’s claims still need to be verified with externally controlled demonstrations.

Tarun Pathak, director of research at Counterpoint Research, said the model’s performance in long-term information, battery consumption during multitasking and reliability across millions of applications 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 consumer processors, said power consumption may be the biggest open question. A model that is capable enough to be used regularly – or continuously in the background in tasks such as an agent – can drain the phone’s battery or require less memory.

What does it mean for the chip demand

The release of PrismML also comes at a time of intense debate over whether improvements in AI efficiency can eventually reduce the need for memory chips and expensive datacenter infrastructure.

Memory has become one of the biggest constraints and costs in all consumer electronics and AI servers. Morgan Stanley estimates that the average cost of Apple’s random access memory could increase by about 190% year-over-year in fiscal 2027, while NAND costs about 180%. NAND is commonly used in flash drives and solid state drives.

The company expects Apple to raise the starting price of the same iPhone 18 models by about $200 to protect margins.

PrismML said its approach could allow cloud models that typically require eight GPUs to run on one, while allowing models that once required a server to run on phones and laptops.

That can reduce the amount of memory or computing power required for a given AI task. But it doesn’t mean that overall chip demand will decline.

Gil Luria, an analyst at DA Davidson, said that shrinking models will not eliminate the need for processors or memory. It can simply move other chips from datacenters to phones and other devices.

“It’s not that you won’t need the chip,” said Luria. “You’ll still need a GPU, and you’ll still need memory.”

He added that using AI on individual devices may be less efficient than using a shared datacenter infrastructure because the chips in phones may sit idle most of the time.

Success can lead to more consumption rather than less spending, as cheap and fast AI powers new products and encourages consumers to use models more often.

Still, the market is quick to chastise anything that suggests AI might need less memory than expected. Micron shares fell in March after Google published its TurboQuant paper on cutting memory usage without harming model performance, although the stock has since recovered.

The public release of PrismML gives everyday users and investors a chance to test whether its claimed benefits stand outside the lab. And for Apple, using more powerful AI directly on the iPhone could help the company improve Siri without giving up the privacy and hardware integration that sets its products apart.

“The combination of cloud and on-device AI can provide a holistic, efficient and privacy-focused AI experience,” said Counterpoint’s Pathak. “Complex tasks will be uploaded to the cloud, and sensitive, sensitive, sensitive and privacy-related tasks will be performed on the device.”

WATCH: Apple is suing OpenAI for allegedly stealing trade secrets: Here’s what you need to know