PrismML’s Compression Breakthrough
PrismML emerged from stealth in March 2026 with what it called the world’s first commercially viable 1-bit large language models, built on research developed at Caltech. The startup’s “Bonsai” model family compresses AI model weights to as little as 1 bit per parameter, shrinking an 8-billion-parameter model from 16 gigabytes to roughly 1 gigabyte while maintaining performance comparable to full-precision models. prismml.com wsj.com prismml.com
Backed by Khosla Ventures, Cerberus Ventures, and Google compute grants, PrismML has since released ternary (1.58-bit) models and an image-generation model that can produce images directly on an iPhone in under 10 seconds. The startup claims to have compressed a 27-billion-parameter AI model to run on an iPhone. binance.com finance.yahoo.com aol.com prismml.com
Vinod Khosla, founder of Khosla Ventures, described PrismML’s technology as “14x smaller, 8x faster” than conventional approaches. x.com
Apple’s On-Device AI Strategy
The discussions come as Apple has been expanding its on-device AI strategy. In May, The Information reported that Apple planned to increasingly showcase the benefits of running models locally on iPhones and other devices rather than in the cloud. Apple has also been active in model efficiency research, presenting related work at ICLR 2026 in April. theinformation.com machinelearning.apple.com
PrismML’s models already run natively on Apple hardware through Apple’s MLX framework, and the startup has released an iOS application called Bonsai Studio that lets users run its image-generation model directly on their phones. Whether the talks between Apple and PrismML could lead to an acquisition, licensing deal, or other arrangement remains unclear from the reporting. finance.yahoo.com progressiverobot.com
