AI startup PrismML has built a small language model for smart glasses powered by Qualcomm chips. At its Snapdragon Summit on September 24, Qualcomm demonstrated PrismML's 1-bit Bonsai model running locally on the Snapdragon AR1 Gen 1 platform, letting wearers ask real-time questions about whatever is in front of them.
The model has 2 billion parameters, tuned for both vision and language tasks. According to TechCrunch, PrismML says it is a quarter the size of larger models while retaining most of their performance on standard benchmarks. The report did not specify which model or which benchmarks were used for the comparison.
PrismML has not announced any commercial glasses shipping with the model yet. Which manufacturer will use it first, and when, remains unclear.
From iPhone to Glasses
PrismML was founded by researchers from Caltech, with Berkeley's Ion Stoica advising the company. Its strategy has consistently been to shrink models small enough to run directly on end devices.
In July, the company said it had compressed Alibaba's Qwen 3.6, a 27-billion-parameter dense model, from roughly 54GB down to under 4GB, running it locally on an iPhone 17 Pro. Its earlier Bonsai series shipped in 8B, 4B, and 1.7B sizes built around 1-bit weights, with the company citing 14 times smaller memory footprint, 8 times faster inference, and 5 times lower energy use compared to full-precision models. Its 1-bit backend, Q1_0, has already been merged into llama.cpp.
This new 2B version sits between Bonsai's existing 1.7B and 4B sizes, with vision capability added on top. Going from 27 billion parameters on an iPhone to 2 billion on glasses, PrismML is pushing toward devices with far less compute headroom.
Why AR1 Gen 1
The Snapdragon AR1 Gen 1 is the chip Qualcomm designed specifically for lightweight smart glasses; display-free AI glasses like Ray-Ban Meta run on this platform. Battery and thermal budgets on glasses are tight, and most AI glasses on the market today hand off visual question-answering to a phone or the cloud, with the glasses themselves only handling photo capture, audio, and playback.
Running the model locally brings two immediate benefits: it doesn't depend on a network connection, so it still works in areas with poor signal, and footage captured by the camera can stay on the device instead of being uploaded. The trade-off is battery life and capability — how complex a question a 2-billion-parameter model can handle, and how much heat and power it draws during continuous recognition, are questions the summit demo couldn't answer; that will have to wait for real hardware.
PrismML's stated goal is to let open-weight AI run on-device, making better use of the compute users already own so they aren't dependent on a handful of large closed-source AI labs.
In China, Qualcomm's AR-series chips are also a common choice for domestic AI glasses makers. PrismML's model follows an open-weight approach, so if it is eventually released for download, Chinese glasses makers and developers could in theory use it directly — but the company has not said when this 2B vision model will be released or under what license.
Sources: TechCrunch, CocoLoop, Qualcomm Snapdragon Summit demo, PrismML's official page; model parameter counts, size ratios, and Bonsai series metrics are per company claims.