PrismML Demonstrates On-Device LLM for Snapdragon Smart Glasses
PrismML model shown on Qualcomm smart-glasses platform
PrismML has adapted its compact language models for smart glasses powered by Qualcomm Snapdragon chips. At Qualcomm’s Snapdragon Summit, the chipmaker demonstrated PrismML’s 1-bit Bonsai LLM running locally on AI glasses based on the Snapdragon AR1 Gen 1 Platform.
The demonstrated model has 2 billion parameters and is tuned for vision and language. Its intended use is real-time interaction with what a wearer is seeing, allowing questions about the surrounding visual scene to be processed on the device.
PrismML was founded by Caltech researchers and is advised by UC Berkeley’s Ion Stoica. The company positions its work around open-weight AI that can run on devices rather than relying entirely on remote proprietary AI services.
A fourfold model reduction claim
PrismML’s central technical claim is that it can substantially reduce larger models while retaining almost all of their performance on standard benchmarks. In this case, the company says its approach shrinks models by a factor of four.
The 1-bit Bonsai LLM demonstration matters because smart glasses have tight constraints around available compute, power and physical space. A smaller model that can execute locally could make it more practical to support vision-and-language queries without sending every interaction to a remote service.
Qualcomm’s Snapdragon AR1 Gen 1 Platform provides the hardware setting for the demo, while PrismML supplies the model optimized for that class of device. The presentation shows a model running on the platform, not a retail product announcement.
No PrismML glasses product announced
No smart glasses carrying PrismML’s model have been announced. The demonstration is instead a step toward the startup’s broader aim: deploying open-weight models on hardware that already has relevant computing capability.
That position is framed as an alternative to depending on the privacy commitments of proprietary AI labs and their demand for increasing compute resources. Local execution can be relevant to organisations assessing where AI processing should occur, especially when a device must respond to visual context in real time.
For businesses evaluating wearable AI, the practical implication is to distinguish platform demonstrations from announced products while tracking whether compact, on-device models can meet the required task performance on available hardware.