September 24, 2026
AI in smart glasses runs locally: Bonsai takes up 0.43 GB on Snapdragon
On Sep 24, PrismML showed 1-bit Bonsai running locally on smart glasses with Snapdragon AR1 Gen 1: the model weights took up 0.43 GB instead of 1.66 GB for the 4-bit version. In a Qualcomm benchmark published by PrismML on Sep 23, 1-bit Bonsai 1.7B generated 15.36 tokens per second, versus 7.44 for the 4-bit model. The image-processing demo combines a 1.7B-parameter language model and a 0.3B image encoder.
On a platform with 4 GB of memory and a 6 TOPS chip, PrismML tested a 1,024-token context. The 1-bit version used 74% less memory, while its speed increased 2.06 times. PrismML optimized the weights and architecture for the Qualcomm Hexagon NPU.
Local start. For publicly available Bonsai models, the docs suggest cloning Bonsai-demo, choosing BONSAI_FAMILY and BONSAI_MODEL, then running `./setup.sh` and `./scripts/run_llama.sh`. These instructions apply to public models, not the VLM from the glasses demo. The available Bonsai 1.7B 1-bit weighs 0.25 GB and is distributed under Apache-2.0.
PrismML aims to bring open-weight models to devices that use already available compute power.
