Apple has found a startup that reduces the size of AI models. This could help it improve Siri

Apple is in talks with PrismML, a small Silicon Valley company that claims it can scale down powerful artificial intelligence models enough to run directly on the iPhone. Babak Hassibi, CEO of the startup PrismML, revealed this in an interview with CNBC.

If the startup’s claims are confirmed by actual testing, the company’s technology could change the demand for memory and computing power in data centers, the TV channel notes, emphasizing, however, that, according to analysts, a large number of chips will still be needed to meet the demands of AI.

Apple shares rose 0.8% in premarket trading on July 15. Year-to-date, they are up nearly 16%.

PrismML uses an open-source language model from Alibaba called Qwen — According to the company’s CEO, the startup has managed to reduce the model’s size from approximately 54 GB to less than 4 GB, making it possible to run all 27 billion of Qwen’s parameters on an iPhone 15 or newer models.

Apple and other companies are currently evaluating the startup’s technology by measuring its speed, energy efficiency, and performance on devices, Hassibi said: “They are indeed looking into our technology right now,” he emphasized, adding that the talks are at a very early stage and their outcome is still unclear.

Apple did not immediately respond to a request for comment.

Apple officially announced the development of PrismML on July 14—the release came a day after Apple launched the public beta testing of iOS 27, giving iPhone owners their first broad access to the long-awaited Siri update, according to CNBC. Apple aims to make Siri more competitive with assistants from OpenAI and Anthropic, while retaining the ability to process data using AI on the device itself.

However, powerful AI models typically require a lot of memory and computing power to run on a smartphone, the TV channel explains. The ability to run them directly on the device would allow Apple to reduce latency associated with accessing remote servers, cut cloud computing costs, and strengthen the company’s case for keeping user data private, explains CNBC. Carolina Milanesi, chief analyst at Creative Strategies, noted that such—more compact AI models—could allow Apple to bring more resource-intensive features to the iPhone—for example, video generation or health-tracking tools that handle sensitive personal data.

"The more features a device has, the better," she said.

The Information had previously reported on PrismML’s technology as a “breakthrough,” citing representatives of the startup: PrismML described its development as “the largest AI model ever launched on an iPhone.” In an interview with CNBC, Hassibi explained that the technology is based on a radical simplification of data storage within the AI model: each value is reduced from 16 bits to one of three possible values. This significantly reduces the amount of memory required to store and run the AI model. Overall, compressed models use 10–15 times less memory, generate responses 6–8 times faster, and consume 3–6 times less energy compared to conventional versions of neural networks running on existing hardware, according to PrismML.

However, there are drawbacks, Hassibi acknowledges: compressed AI models typically lose a few percentage points in overall performance. Accuracy in reproducing facts suffers the most, whereas reasoning skills, the ability to perform mathematical calculations, and programming abilities are preserved better.

PrismML has released two lightweight versions of the model for free, according to CNBC. They are designed to run on standard devices, including iPhones, MacBooks, and PCs with Nvidia graphics cards.

PrismML was founded in 2025. Its technology, as CNBC explains, grew out of Hassibi’s research group at the California Institute of Technology. He holds the patents underlying the development and grants PrismML exclusive licenses to use them. In March, the company raised $16.25 million in a seed funding round backed by venture capital firm Khosla Ventures and other investors.

Horace Dediu, founder of the research firm Asymco, noted that Apple will likely try to keep the vast majority of common interactions with Siri on the device, reserving the most resource-intensive tasks for the cloud. In his view, the advantage of using compressed neural networks lies not simply in using less memory, but in the ability to fit a more powerful model within the same physical constraints. “They [Apple] are trying to figure out how large and how ‘smart’ a model they can fit onto the device,” said Dediu (as quoted by CNBC).

However, PrismML’s claims still need to be validated outside of controlled demonstrations, notes Tarun Pathak, research director at Counterpoint Research: “The real test will be millions of requests, thousands of combinations of [different] devices, and thorough testing on an enterprise-wide scale,” he added.

"Such AI models—which are powerful enough for frequent use—or for continuous operation in the background to perform agent-like tasks," noted Phil Solis, head of client processor research at IDC. He believes that using such a neural network could drain a phone’s battery faster, even if it requires less memory.

DA Davidson analyst Gil Luria agreed that running AI models on individual devices may be less efficient than using a shared data center infrastructure; he suggested that chips in phones might sit idle most of the time. Furthermore, the expert believes that compressing AI models could lead to a significant number of chips being moved from data centers to smartphones and other devices, but it won’t eliminate the need for processors or memory chips entirely: “It’s not that you won’t need a chip. You’ll still need a graphics processor and memory,” Luria said.

This article was AI-translated and verified by a human editor