PrismML is pushing micro-language models into the smart glasses scenario. Qualcomm demonstrated on Snapdragon Summit that the 1-bit Bonsai LLM developed by this company can now run locally on smart glasses based on the Snapdragon AR1 Gen 1 platform.

This means that some visual and language tasks can be completed directly on the device side, without the need to rely on cloud processing. According to the settings of PrismML, the wearer can ask questions in real-time about what they see, and the device will provide corresponding answers.

2B parameter version is oriented towards visual language.

For this application in smart glasses, a model with 2 billion parameters was used, which has been optimized for visual and language tasks. The article states that the main selling point of PrismML is the ability to compress a larger model into a much smaller size; in this case, it can be reduced by about 4 times while still maintaining most of the performance seen in standard benchmark tests.

PrismML was founded by researchers from the California Institute of Technology, and its advisory team includes professors from the University of California, Berkeley. The company aims to promote the operation of open-weight models on mobile devices, making more efficient use of the existing hardware computing power.

On the client side, AI focuses on privacy and computing efficiency.

PrismML positions this approach as an alternative, which is to reduce reliance on closed AI laboratories. The idea is to perform more inference tasks locally, thereby lowering the dependence on cloud computing power and external privacy commitments.

For Qualcomm, such models also help to enhance the edge-side capabilities of Snapdragon in new hardware such as AI glasses. As wearable devices incorporate more real-time recognition and question-answering functions, the size, power consumption, and response speed of the models are becoming increasingly crucial.

Products to be equipped with this technology have not yet been announced.

However, there are currently no smart glasses equipped with PrismML officially released. The focus of the information disclosed at this stage is still that the model has completed adaptation to Qualcomm's chip platform and was demonstrated at the conference.

This takes PrismML one step closer to its goal of “opening up device-side weight AI”, but whether it can truly enter the consumer-grade hardware market still depends on whether manufacturers will launch corresponding products in the future.