PrismML debuts 1-bit LLM for Qualcomm Snapdragon AR glasses

PrismML used Qualcomm’s Snapdragon Summit on Wednesday to demonstrate its latest “Bonsai” language model, a 1-bit architecture that can operate directly on smart-glasses powered by the Snapdragon AR1 Gen 1 platform. By fitting the model onto the edge device, the company promises real-time visual-language queries without relying on cloud services, a move that could reshape how wearable AI handles user data.

The model presented is a 2-billion-parameter network that has been compressed to one-quarter of the size of comparable large language models while preserving nearly the same benchmark scores. PrismML’s technique, which it describes as a four-fold reduction in parameters, enables the model to run locally on the limited compute budget of AR glasses, delivering instant answers about objects in the wearer’s field of view.

Founded by researchers from Caltech and guided by UC Berkeley professor Ion Stoica, PrismML has positioned itself as a champion of open-weight, on-device artificial intelligence. The startup’s broader ambition is to provide alternatives to the proprietary AI offerings that dominate the market and that typically require continuous data transmission to powerful data centers. By leveraging the existing processing power of devices like Qualcomm’s AR chips, PrismML aims to reduce dependence on external compute while enhancing user privacy.

For founders and investors, the announcement signals a tangible step toward commercially viable edge AI that does not sacrifice performance. If the technology can be integrated into consumer-grade smart glasses, it could open new revenue streams for hardware manufacturers and create a market for applications that need immediate, private inference. The ability to run sophisticated language-vision models on a wearable also lowers barriers for developers seeking to embed AI into niche use cases such as field service, medical assistance, or industrial inspection.

While the model is now available for Qualcomm’s platform, PrismML has not yet disclosed a partner or product that will ship with the technology. The company’s next milestone will likely involve a hardware collaboration that brings the Bonsai LLM to market. Observers will watch to see whether the promise of on-device, privacy-preserving AI can translate into a viable consumer product and how quickly the ecosystem around Qualcomm’s AR hardware can adopt it.