PrismML says it compressed a 27-billion-parameter model from 54GB to under 4GB, letting it run on an iPhone 15 or newer
CFOTO/Future Publishing via Getty Images
PrismML released its Bonsai 27B model on Tuesday, which the company said reduces a 27.8-billion-parameter model from roughly 54GB to as little as 3.9GB by storing each model weight as one or three possible values rather than a standard 16-bit figure. At that size, the model can run on an iPhone 15 or newer, the company said.
Hassibi told CNBC that Apple and other companies have been measuring the startup's models for speed, energy use, and on-device performance. Apple didn't comment.
According to PrismML, the compressed models require one-tenth to one-fifteenth the memory of standard versions, deliver responses at six to eight times the speed, and draw three to six times less power. Hassibi noted that compression comes at a cost: the models shed a small number of percentage points in overall accuracy, and factual knowledge degrades sooner than capabilities like reasoning and coding.
The Bonsai 27B series is available in two variants. The 1-bit version, at roughly 4GB, is optimized for maximum compression and is designed to run on high-end mobile devices including the iPhone 17 Pro, where the company said it achieves 11 tokens per second. The 1.58-bit ternary version offers higher quality while remaining significantly smaller than a full-precision model, and achieves more than 95% of full-precision benchmark performance, the company said. Both models are available under a free Apache 2.0 license. PrismML said Google $GOOGL's open-source Gemma model is next in its pipeline.
PrismML was spun out of research conducted at the California Institute of Technology, which holds the relevant patents and grants PrismML an exclusive license to commercialize them. The startup raised a $16.25 million seed round in March backed by Khosla Ventures.
Apple has been working to expand how much AI processing happens directly on its devices rather than in the cloud. Apple unveiled an overhauled version of Siri at its Worldwide Developers Conference in June, powered by Google Gemini, with iOS 27 entering public beta on Monday. Apple presently sends more demanding queries to its private cloud servers or third-party models, reserving a subset of tasks for on-device processing. Shifting more AI work onto the iPhone itself would cut response times, trim expenditures on cloud infrastructure, and reinforce Apple's long-standing privacy narrative.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
