PrismML says it ran a 2-billion-parameter vision and language model locally on smart glasses hardware built around Qualcomm’s Snapdragon AR1 Gen 1 platform. The company demonstrated the model at Snapdragon Summit. TechCrunch reports that no glasses using it have been announced for sale.
The demo uses a 1.7-billion-parameter language model with 1-bit weights and a 0.3-billion-parameter vision encoder at 4-bit precision. PrismML says this setup can interpret what a wearer sees and answer questions on the device. That is a useful direction for glasses, where memory, heat, and battery capacity limit how much processing can happen locally.
What the published numbers measure
PrismML’s announcement includes details of Qualcomm’s September test. On a Snapdragon AR1 Gen 1 configuration with 4 GB of memory and a 1,024-token context, the 1-bit language model used 0.43 GB for its weights. A corresponding 4-bit version used 1.66 GB. That is a 74% reduction in weight memory, or about 3.8 times less.
Qualcomm also measured 15.36 generated tokens per second for the 1-bit model, compared with 7.44 for the 4-bit version on the same platform. The speed difference was about twofold in that test. PrismML separately says its benchmark scores are comparable to a 4-bit Qwen 3 model of the same language-model size, based on its own evaluations.
These results describe a specific model, compiler, chip configuration, and test method. The release does not give battery life, heat, or response quality in a finished pair of glasses. It also does not establish how well the model handles a busy camera scene or a longer conversation beyond the tested context.
