Is Apple's (AAPL.US) AI dilemma solved? Negotiate AI startup Prismml to “slim down” the big model into the iPhone
The Zhitong Finance App learned that Apple (AAPL.US) is in talks with PrismML, a small startup in Silicon Valley. The latter claims to be able to drastically compress the powerful AI model, making it enough to run directly on the iPhone.
Technological breakthrough: 54GB model reduced to less than 4GB, 27 billion parameters landed on iPhone
PRISMML, a Caltech (Caltech) spinoff with investment from Khosla Ventures, publicly released a compressed version of Alibaba's open source model Qwen on Tuesday. The company said it has reduced the original model from about 54GB to less than 4GB, so that all 27 billion parameters can run on the iPhone 15 and newer models.
PrisMML CEO Babak Hassibi (Babak Hassibi) revealed that Apple and others are evaluating the startup's model to test its operating speed, energy efficiency, and performance on the device side.
“They're really evaluating our technology right now,” Hasibie said of Apple. He described the relevant discussions as still in the “very early stages”, and the final direction is still unclear, but “things are progressing steadily.”
Apple's AI Dilemma: Trade-off and Balance Techniques for End-side Intelligence
The release of the PrisMML model coincided with Apple officially opening the iOS 27 public beta the day before, and iPhone users were able to experience the full Siri upgrade after a long delay on a large scale for the first time. Apple is trying to make Siri more competitive in competition with OpenAI and Anthropic, while insisting on leaving more personal information and AI processing on the device side.
However, the most powerful AI models usually require huge amounts of memory and computing power to run on smartphones, which forms a core contradiction in Apple's AI strategy. Apple's current approach is to send complex requests to the cloud model for processing, but running more AI capabilities directly on the iPhone can reduce data transmission delays, reduce cloud computing costs, strengthen privacy protection commitments, and enable some functions to work properly without an internet connection.
Carolina Milanesi (Carolina Milanesi), president and chief analyst at Creative Strategies, points out that the smaller model will allow Apple to move more challenging features such as computational photography, video generation, and health or fitness tools that rely on sensitive personal data to run locally on the iPhone. “The more things can be done on the device, the better,” she said, citing health and medication data as an example. Users want this information to remain private.
16 bits reduced to “1 or 3 values”, and the speed soared 6 to 8 times
According to PrisMML, its compression technology is implemented by greatly simplifying the storage method of model internal information — reducing each value from 16 bits to only 1 or 3 possible values, thereby significantly reducing the memory required for model storage and operation.
Hasibie likened this breakthrough to the chip industry's evolution from 8-bit to 4-bit computing, but it went a step further. Allegedly, the memory usage of the compressed model is only one-tenth to one-fifteenth of the traditional version, the response speed is 6 to 8 times faster, and the energy consumption is reduced by 3 to 6 times.
However, Hasibie admits that there is a loss in performance. PrisMML models usually lose a few percentage points in overall performance, with factual memory being weakened more than ability to reason, math, and coding.
PrisMML released two compressed models, pointing to robots and autonomous driving
Prismml is releasing two free compressed models. The design goal is to run on everyday devices, including iPhones, MacBooks, and PCs equipped with Nvidia chips.
The technology was born out of Hasibe's research team at Caltech (Caltech). The school holds the underlying patent and is exclusively authorized for use by PrismML. In March of this year, the company completed a $16.25 million seed round led by Khosla Ventures and followed by various institutions.
Hasibie revealed that Google's open source model Gemma will be the next compression target, and will challenge larger models thereafter — including the cutting-edge model of Head Lab, which currently usually requires data center hardware to operate.
According to PrisMML, the technology's application scenarios will eventually extend far beyond mobile phones and laptops to robots, autonomous driving systems, and other products that require quick decision-making without relying on cloud connectivity.
“Intelligence has to be localized and it has to run fast, which is critical,” Hasibie said.
Apple has an edge on the end side, and scale tests and battery life tests need to be solved
Apple already runs some AI features locally on its devices, including translation, some summaries, and features closely related to personal information. More complex requests are routed to Apple's private cloud infrastructure or external third-party models.
Horace Dediu (Horace Dediu), founder of Asymco, said that Apple is likely trying to keep the vast majority of common Siri interactions on the device and leave only the heaviest tasks to the cloud.
He pointed out that the advantage is not only to reduce memory usage, but to include a more capable model within the same physical limitations.
“They're trying to figure out how big and smart models they can fit into the device,” Dediu said. Leaving common requests processed locally can bring Apple lower latency, greater privacy protection, and potentially lower licensing fees and cloud service costs.
Because of its self-designed iPhone chip and software, and the integration of software and hardware makes it have more precise control over the operation of AI on the device side, Apple may have a unique advantage in implementing these models.
Analysts are generally cautiously optimistic about PrisMML's technology, but emphasized that it still needs to be tested outside of controlled demonstrations.
Counterpoint Research's research director Tarun Pathak (Tarun Pathak) pointed out that the model's performance when processing long prompts, battery consumption in multi-tasking scenarios, and reliability in dealing with millions of queries will be key. “The ultimate test will be millions of queries, thousands of device combinations, and large-scale robust testing,” Patak said.
Phil Solis (Phil Solis), head of client processor research at IDC, believes that power consumption is probably the biggest unknown. Even if the model requires less memory, if it is capable enough to be used frequently or run continuously in the background (such as proxy tasks), it may still consume a significant amount of battery power on the phone.
Chip Demand Dispute: Increased Efficiency ≠ Decreased Demand
When PrisMML released its compression model, the market was fiercely debating whether improving AI efficiency would ultimately reduce the need for memory chips and expensive data center infrastructure.
Storage has become one of the biggest bottlenecks and costs in consumer electronics and AI servers. Morgan Stanley predicts that Apple's average DRAM cost per bit for the 2027 fiscal year may increase by about 190% year on year, and NAND costs will rise by about 180%. The agency expects Apple to increase the starting price of the iPhone 18 series models by about 200 US dollars to protect profit margins.
PrisMML said that its technology enables cloud models that originally required 8 GPUs to run with only 1, while also enabling models that originally relied on servers to migrate to mobile phones and laptops.
But D.A. Davidson analyst Gil Luria (Gil Luria) points out that model compression won't eliminate the need for processors or storage; it simply moves more chips from data centers to phones and other devices. “It's not that you don't need a chip anymore; you still need a GPU, you still need memory,” Luria said. He also added that running AI on personal devices may actually not be as efficient as sharing data center infrastructure because the chips in the phone may be idle most of the time.
Furthermore, efficiency breakthroughs often lead to more usage rather than lower expenses — cheaper and faster AI will spawn new products and drive users to run models more frequently.
The public version of PrisMML has been released, providing regular users and investors with an opportunity to verify its performance improvements outside of the lab. As far as Apple is concerned, the AI model with more ability to run on the iPhone will help achieve a substantial upgrade of Siri without abandoning the advantages of privacy protection and software integration.
Counterpoint's Patak concluded, “The combination of cloud and end-side AI can provide a more complete, more efficient, and privacy-focused AI experience. Complex tasks will be offloaded to the cloud, while sensitive, delay-sensitive, and privacy-related tasks will be executed on the device side.”