PrismML Shrinks Its 1-Bit LLM to Fit on Qualcomm Smart Glasses

Qualcomm showed a 1-bit language model from startup PrismML running locally on AI smart glasses at its Snapdragon Summit on Wednesday. The demonstration puts a 2-billion-parameter model, tuned for vision and language, on hardware built around Qualcomm's Snapdragon AR1 Gen 1 Platform.

The model, called Bonsai, comes from AI Lab PrismML, a startup founded by Caltech researchers and advised by UC Berkeley's Ion Stoica. It lets glasses wearers ask what they are looking at in real time.

The technical claim behind the demo is compression. PrismML says its 1-bit approach shrinks larger models by 4x while retaining almost all of their performance on standard benchmarks, as TechCrunch previously reported. That compression ratio is what makes local execution feasible on a power-constrained wearable chip rather than a data-center GPU.

PrismML's larger ambition is open-weight AI that runs directly on consumer devices and squeezes more out of the computing power those devices already carry. The startup frames this as an alternative to relying on the privacy promises of proprietary AI labs — and to those labs' insatiable demand for more compute.

Shipping a model optimized for Qualcomm's silicon moves PrismML toward that vision. But the gap between a summit demo and a shipping product remains: no smart glasses running PrismML have been announced yet.

The demonstration still signals where on-device AI may be heading. If 1-bit models can deliver near-benchmark performance at a quarter of the size, wearable makers gain a path to real-time, offline AI features without paying cloud inference costs or shipping user data off-device.