Apple Eyes PrismML to Shrink AI Models for iPhones
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Apple is evaluating PrismML's compression tech; the startup says it shrunk Alibaba's Qwen from roughly 54 GB to under 4 GB, letting all 27 billion parameters run on an iPhone 15 or newer.
- PrismML cuts memory by reducing each stored value from 16 bits to one or three possible values, claiming 10-15x lower memory use, 6-8x faster responses, and 3-6x lower energy consumption than un-compressed versions.
- The release landed a day after iOS 27's public beta opened, giving iPhone owners their first broad access to Apple's long-delayed Siri overhaul aimed at matching assistants from OpenAI and Anthropic.
- PrismML, a Caltech spinout that raised a $16.25 million Khosla Ventures seed in March and exclusively licenses Caltech's underlying patents, released two compressed Qwen variants free and plans to compress Google's Gemma next.
- Morgan Stanley projects Apple's DRAM cost per bit will rise roughly 190% year-over-year in fiscal 2027 and NAND costs about 180%, and expects the iPhone 18 starting price to climb roughly $200 to protect margins.
Why it matters: On-device AI has been the binding limit on Apple's Siri strategy — every query routed to the cloud costs latency, licensing fees, and privacy credibility. If PrismML's 10-15x memory reduction survives real-world testing, Apple shifts routine Siri workloads onto the device itself and trims its dependence on cloud pacts with OpenAI and Anthropic — directly relevant as Morgan Stanley pegs Apple's DRAM cost-per-bit up ~190% in fiscal 2027 and projects a ~$200 iPhone 18 starting-price hike to offset the squeeze.
