Apple is reportedly in very early discussions with PrismML, a Silicon Valley startup spun out of Caltech and backed by Khosla Ventures, about technology that could significantly reduce the size of powerful AI models and run them directly on the iPhone. CNBC reported.
PrismML recently announced that it has compressed Alibaba’s open source 27 billion parameter Qwen model from approximately 54 GB to less than 4 GB, making all 27 billion parameters run on iPhone 15 and later. The company says its compression method reduces memory usage by 10 to 15 times, improves processing speed by 6 to 8 times, and reduces energy consumption by 3 to 6 times, but the overall performance sacrifice is only a few percent, and the reduction in fact recall is greater than in reasoning, math, or coding ability.
Apple is reportedly evaluating the technology’s speed, energy efficiency, and on-device performance as part of a broader effort to improve Siri and expand its on-device AI capabilities. Running more AI locally enables faster responses, increases user privacy by keeping more sensitive data on-device, reduces reliance on cloud computing, and allows certain AI capabilities to operate without an internet connection.
Analysts say this approach could support more advanced applications such as computational photography, video generation, and health and fitness tools that process personal information. However, they caution that PrismML’s claims need to be verified through extensive real-world testing, especially when it comes to battery life, reliability, and performance across millions of devices.
If the technology works as claimed, it could help Apple advance its AI strategy while increasing its focus on privacy and hardware-software integration. Although this breakthrough could shift a larger share of AI processing from data centers to personal devices, experts say it is unlikely to reduce overall demand for processors and memory chips.
CNBC has the full story.
