PrismML has developed a specialized version of its Bonsai language model, designed to run locally on smart glasses powered by Qualcomm Snapdragon chips. The 1-bit model was showcased at the Snapdragon Summit, running on the Snapdragon AR1 Gen 1 platform.
The model features 2 billion parameters and is optimized to process both text and visual data. PrismML achieved a fourfold compression of the model while maintaining quality comparable to 4-bit systems. This allows users to receive real-time information about their surroundings directly through the glasses.
The model is specifically optimized for the Hexagon neural processor. According to the developers, it consumes four times less memory and generates tokens more than twice as fast as previous iterations. This approach aims to reduce reliance on cloud-based proprietary AI services and improve data privacy.
PrismML provides models with open weights to maximize the use of local device computing power. The company plans to continue its partnership with Qualcomm Technologies to further optimize future versions of the Bonsai model for the Snapdragon platform.
