PrismML, an AI lab founded by Caltech researchers and advised by UC Berkeley’s Ion Stoica, has developed compact language models designed to run on smart glasses powered by Qualcomm’s Snapdragon chips. This week at Qualcomm’s Snapdragon Summit, the company demonstrated its 1-bit Bonsai large language model (LLM) operating locally on AI smart glasses using the Snapdragon AR1 Gen 1 Platform. PrismML’s approach significantly reduces model size by a factor of four while maintaining near-original performance on standard benchmarks.

The smart glasses iteration employs a 2-billion-parameter model fine-tuned for both vision and language tasks, enabling users to inquire about their surroundings and receive real-time responses. This integration marks a step toward PrismML’s broader mission to create open-weight AI systems capable of operating efficiently on edge devices. By leveraging existing device computing power, PrismML aims to minimize reliance on centralized, proprietary AI systems, which often raise privacy concerns and demand substantial computational resources.

PrismML positions its technology as a privacy-conscious alternative to mainstream AI models that require cloud-based data transmission and heavy compute infrastructure. The ability to run LLMs locally on devices like smart glasses reduces exposure of sensitive data and can boost responsiveness and energy efficiency. This technology aligns with trends toward decentralized AI, where machine learning capabilities can be embedded into smaller, wearable hardware with constrained processing power.

While Qualcomm’s Snapdragon-powered smart glasses featuring PrismML’s tiny LLMs were showcased at the event, no specific consumer devices incorporating this technology have been announced yet. The demonstration underscores the potential for future AR glasses to offer sophisticated, AI-powered interactions with the environment, assisted by compact but capable language models running directly on the headset hardware.