PrismML, an AI lab founded by Caltech researchers and advised by UC Berkeley’s Ion Stoica, has developed a small language-and-vision model intended for smart glasses that run on Qualcomm Snapdragon chips. The announcement was made at Qualcomm’s Snapdragon Summit on Wednesday, where Qualcomm showcased PrismML’s 1-bit Bonsai LLM.
What the model is
- The version presented is a 2-billion-parameter model tuned for vision and language tasks, allowing wearers to ask about what they are seeing in real time.
- PrismML says the Bonsai variant compresses larger models by about 4x while retaining almost all of their performance on standard benchmarks.
- The model uses 1-bit quantization and can run locally on devices built on the Qualcomm Snapdragon AR1 Gen 1 Platform.
Why this matters
PrismML’s broader goal is to promote open-weight AI that runs on devices and better uses their existing compute, reducing reliance on cloud services. The startup positions this approach as an alternative to depending on proprietary AI labs and their growing compute demands, with implications for on-device privacy and latency.
Current status and next steps
Releasing a model adapted to Qualcomm’s chip is a concrete step toward PrismML’s vision, but no commercial smart glasses running PrismML have been announced to date. The demonstration indicates progress in the software and chip-level support, while actual end-user devices incorporating the model remain to be revealed.
Summary
PrismML’s 2-billion-parameter, vision-and-language Bonsai model runs locally on the Snapdragon AR1 Gen 1 Platform with 1-bit quantization. The company claims roughly 4x size reduction with minimal loss in benchmark performance, advancing its aim of on-device, open-weight AI — though hardware partners or shipping smart glasses have not yet been disclosed.
