PrismML's Compressed Models Land on Qualcomm Smart Glasses

Researchers and industry players are racing to put capable AI directly onto edge devices—and PrismML, a startup founded by Caltech researchers and backed by UC Berkeley's Ion Stoica, just took a significant step in that direction. The team unveiled an optimized version of its compressed language models specifically for Qualcomm's latest smart glasses platform at the Snapdragon Summit this week.

What's New#

The centerpiece is a 2-billion-parameter variant of PrismML's Bonsai LLM, tuned to handle both vision and language tasks. Wearers can ask questions about their surroundings in real time, with the model running entirely on the Snapdragon AR1 Gen 1 Platform without needing to phone home to a remote server.

The Compression Breakthrough#

PrismML's core innovation is extreme model compression: it shrinks existing language models by roughly 4x while maintaining comparable performance on standard benchmarks. By converting models to 1-bit precision, the startup strips away computational overhead without sacrificing core capabilities—a feat that lets cutting-edge AI fit onto silicon with far less power and memory than traditional approaches demand.

The Larger Vision#

The startup frames this as a privacy-first and efficiency-focused alternative to the current cloud-centric AI model. Rather than relying on the data-handling policies of proprietary AI labs and their endless hunger for computational resources, PrismML champions open-weight models that run locally on consumer hardware, letting devices make full use of their existing silicon.

Moving onto Qualcomm's chips represents meaningful progress toward that goal—though no commercial smart glasses shipping with PrismML's models have been announced yet.