Reports indicate that Apple (AAPL.US) is in discussions with PrismML, a small Silicon Valley startup. The company claims its technology can significantly compress powerful AI models, making them small enough to run directly on an iPhone.
Key Technical Achievement
PrismML, a Caltech spin-off backed by Khosla Ventures, publicly released a compressed version of Alibaba's open-source Qwen model on Tuesday. The company states it has reduced the original model's size from approximately 54GB to under 4GB, enabling all 27 billion parameters to operate on an iPhone 15 and newer devices. PrismML CEO Babak Hassibi confirmed that Apple and other companies are currently evaluating the startup's model, testing its on-device speed, energy efficiency, and performance. Hassibi described discussions with Apple as being at a "very early" stage, with the ultimate outcome uncertain, but noted that "things are moving along."
Apple's AI Strategy Challenge
The timing of PrismML's release follows Apple's public beta launch of iOS 27, which allows iPhone users to experience a long-delayed, comprehensive Siri upgrade. Apple aims to make Siri more competitive against offerings from OpenAI and Anthropic while maintaining its commitment to keeping more personal data and AI processing on the device itself. However, the most powerful AI models typically require substantial memory and computing power to run on smartphones, creating a core tension in Apple's AI strategy. Apple's current approach involves sending complex requests to cloud-based models for processing, but moving more AI capabilities directly onto the iPhone can reduce data transfer latency, lower cloud computing costs, strengthen privacy promises, and allow some features to function without an internet connection. Carolina Milanesi, President and Principal Analyst at Creative Strategies, noted that smaller models would enable Apple to shift more challenging functions like computational photography, video generation, and health or fitness tools reliant on sensitive personal data to run locally on the iPhone. "The more you can do on the device, the better," she stated, using health and medication data as an example of information users want to keep private.
Compression Technique Details
PrismML explains that its compression technology works by drastically simplifying how information is stored within the model—reducing each numerical value from 16 bits down to just 1 or 3 possible values. This significantly cuts the memory required to store and run the model. Hassibi likened this breakthrough to the chip industry's evolution from 8-bit to 4-bit computing, but taken a step further. The compressed model reportedly uses one-tenth to one-fifteenth the memory of traditional versions, with response speeds 6 to 8 times faster and energy consumption 3 to 6 times lower. However, Hassibi acknowledged a trade-off in performance. PrismML's models typically see a performance drop of a few percentage points overall, with a greater reduction in factual recall capability compared to skills like reasoning, math, and coding.
Broader Applications and Goals
PrismML has released two compressed models for free, designed to run on everyday devices including iPhones, MacBooks, and PCs with NVIDIA chips. The technology originated from Hassibi's research team at Caltech. The university holds the underlying patents and has granted an exclusive license to PrismML. In March, the company closed a $16.25 million seed funding round led by Khosla Ventures. Hassibi revealed that Google's open-source Gemma model is the next target for compression, followed by attempts on larger models, including cutting-edge models from top labs that currently require data center hardware to run. According to PrismML, the technology's applications could eventually extend far beyond phones and laptops to robots, autonomous driving systems, and other products requiring rapid decision-making without relying on a cloud connection. "The intelligence has to be local, and it has to be fast. That's critical," Hassibi said.
Apple's Potential Edge and Remaining Hurdles
Apple already runs some AI features locally on its devices, including translation, some summarization, and features closely tied to personal information. More complex requests are routed to Apple's private cloud infrastructure or external third-party models. Horace Dediu, founder of Asymco, suggested Apple is likely trying to keep the vast majority of common Siri interactions on the device, sending only the most demanding tasks to the cloud. He noted the advantage is not just about using less memory, but fitting a more capable model within the same physical constraints. "They are trying to figure out how big and how smart a model they can stuff on the device," Dediu said. Processing common requests locally offers Apple benefits like lower latency, stronger privacy, and potential reductions in licensing fees and cloud service costs. Because Apple designs its own iPhone chips and software, this vertical integration may give it a unique advantage in fine-tuning how these models run on-device. Analysts are cautiously optimistic about PrismML's technology but emphasize it needs testing beyond controlled demonstrations. Tarun Pathak, Research Director at Counterpoint Research, pointed out that key tests will involve the model's performance with long prompts, battery consumption during multi-tasking, and reliability across millions of queries. "The ultimate test will be millions of queries, thousands of device combinations, and large-scale robustness testing," Pathak stated. Phil Solis, Research Director for Client Processors at IDC, believes power consumption may be the biggest unknown. Even if a model requires less memory, if it's capable enough to be used frequently or run persistently in the background for agent-like tasks, it could still significantly drain a phone's battery.
Impact on Chip Demand
PrismML's release comes amid a broader market debate over whether AI efficiency gains will ultimately reduce demand for memory chips and expensive data center infrastructure. Memory has become one of the biggest bottlenecks and cost items in both consumer electronics and AI servers. Morgan Stanley estimates that Apple's average per-bit DRAM cost could rise about 190% year-over-year in fiscal 2027, with NAND costs up about 180%. The firm expects Apple to raise the starting price of the iPhone 18 series by approximately $200 for equivalent configurations to protect margins. PrismML claims its technology could allow a cloud model that previously required 8 GPUs to run on just 1, and also enable models that once needed servers to migrate to phones and laptops. However, D.A. Davidson analyst Gil Luria noted that model compression does not eliminate the need for processors or memory; it simply shifts more chips from data centers to phones and other devices. "It's not that you don't need chips. You still need GPUs, you still need memory," Luria said. He added that running AI on personal devices might actually be less efficient than shared data center infrastructure, as chips in phones could be idle much of the time. Furthermore, efficiency breakthroughs often lead to more usage, not lower spending—cheaper, faster AI will spawn new products and encourage users to run models more frequently.
Conclusion
The public release of PrismML's models provides an opportunity for users and investors to verify its performance claims outside a lab setting. For Apple, the ability to run more capable AI models on the iPhone could enable a substantial upgrade for Siri without abandoning its commitments to privacy and the advantages of its integrated hardware and software. As Counterpoint's Tarun Pathak summarized, "The combination of cloud and on-device AI can deliver a more complete, efficient, and privacy-focused AI experience. Complex tasks will be offloaded to the cloud, while sensitive, latency-sensitive, and privacy-involving tasks will be executed on the device."
Comments