Qualcomm and PrismML Run 1-Bit Bonsai Vision Model On-Device on Snapdragon AR1 Gen 1

At the 2026 Snapdragon Summit, Qualcomm and PrismML announced they have successfully run a 2-billion-parameter 1-bit quantized Bonsai vision model locally on the Snapdragon AR1 Gen 1 smart glasses platform. The model is built on PrismML's previously released Bonsai 1.7B and compresses model precision to 1-bit, reducing the memory required for storage and inference.

According to the companies, the 1-bit Bonsai model uses about one-quarter of the memory of a traditional 4-bit quantized model, fits roughly four times more parameters within the same memory capacity, and doubles text token generation speed. This allows use cases such as identifying objects in view, translating text, answering questions about the environment, and providing hands-free voice assistance to run directly on the device without cloud connectivity.

The lower memory and power requirements are also expected to help extend battery life and reduce heat, improving the overall wearing experience. The collaboration highlights a path for deploying on-device AI in lightweight smart hardware.