At the Snapdragon Summit 2026, Qualcomm teamed up with PrismML to locally execute a 2-billion-parameter, 1-bit quantized Bonsai vision model on the Snapdragon AR1 Gen 1 smart glasses platform. This model, based on Bonsai 1.7B, reduces memory usage to merely one-fourth of that of a 4-bit model by employing 1-bit precision compression. Consequently, it can accommodate roughly four times the parameters within the same memory capacity. This breakthrough substantially diminishes the demands for storage and inference resources, thereby facilitating offline image recognition and question-and-answer functionalities.
