A Caltech-rooted startup called PrismML released a compressed AI model on July 14 capable of running locally on an Apple (NASDAQ: AAPL) iPhone 17 Pro without requiring a cloud connection.
The model, called Bonsai 27B, is built on Alibaba’s open-weight Qwen3.6 architecture and has been compressed to just 3.9 gigabytes while retaining roughly 90% of the original model’s performance.
A traditional 27-billion-parameter model cannot fit within a smartphone’s usable memory, making Bonsai 27B a meaningful technical achievement in the field of model compression.
Beyond the iPhone 17 Pro, PrismML’s model is also compatible with iPad, Mac, and PC devices, broadening its potential reach across Apple’s entire hardware ecosystem.
PrismML’s CEO recently told CNBC that Apple is in “very early” discussions about the technology, signaling that a formal commercial relationship between the two companies could eventually take shape.
For Apple, a local AI wave presents a mixed picture, as more capable on-device models would require greater memory and processing capacity, raising the cost of goods sold for its hardware business.
Apple’s own AI execution has faced headwinds, with the previously delayed and rebuilt Siri falling short in internal testing as recently as February before entering public beta testing weeks ago.
Management sounded optimistic during Apple’s July 30 earnings call, noting that early feedback from Siri AI’s public beta has been positive, suggesting the product may be recovering momentum.
Apple is also engaged in a legal battle after suing OpenAI in federal court in July, alleging trade-secret theft tied to former engineers who joined the lab and helped OpenAI build its own devices.
The company stayed silent about the lawsuit during its earnings call, though the case places Apple squarely at the center of a widening fight over the future of consumer AI hardware and data.
With breakthroughs like Bonsai 27B emerging from independent developers, consumers may be approaching an era of capable, offline AI that requires no monthly subscription and keeps personal data stored on the device.
Apple’s custom silicon was designed with exactly this kind of workload in mind, positioning the company well if the broader market shifts toward locally run, open-weight AI models.
