PrismML Brings Compact LLMs to Qualcomm-Powered Smart Glasses

PrismML announced on September 23, 2026 that its 1-bit Bonsai model can now run locally on AI smart glasses powered by Qualcomm's Snapdragon AR1 Gen 1 Platform .

The model is a 2-billion-parameter vision-language model built on PrismML's Bonsai 1.7B . It was demonstrated at the Snapdragon Summit running on the glasses hardware itself, with no cloud connection needed .

Key technical figures from the announcement:

- 4x more parameters fit into the same memory footprint compared to 4-bit models

- ~4x less memory used, with the language model weight memory dropping from 1.66 GB to 0.43 GB

- 2x faster token generation, from 7.44 to 15.36 tokens per second

PrismML worked with Qualcomm to optimize the model's weights and architecture for the Hexagon NPU, taking advantage of the hardware to maximize performance and efficiency . The company describes this as model-hardware co-design, arguing that future AI progress cannot come from adding more compute alone but from getting dramatically more intelligence from the compute already present .

The model enables experiences such as understanding what the wearer is seeing, reasoning over that context, and responding in real time . For example, a wearer could look at an object and ask about it, with the glasses answering locally without a cloud round-trip .

PrismML's wider goal is open-weight AI that runs directly on devices and makes better use of the computing power they already have . The startup positions this as an alternative to depending on the privacy promises of proprietary AI labs and their growing need for more compute . PrismML emerged from Caltech research and is backed by Khosla Ventures, Cerberus, and Google, with support from Samsung, Gates Frontier, and Jeff Dean .