PrismML Brings 1-Bit Bonsai LLM to Snapdragon Smart Glasses

The AI industry has mostly measured progress by size: more parameters, more data centres, more compute. That makes sense for a frontier lab. It makes less sense for a pair of glasses that has to answer a simple question: what am I looking at right now?

PrismML is building for the second case. The AI lab was founded by researchers from Caltech and counts UC Berkeley's Ion Stoica among its advisers. It has now produced a version of its compact language models for smart glasses that run on Qualcomm's Snapdragon chips.

From the data centre to the frame: what Qualcomm showed at Snapdragon Summit

On Wednesday, Qualcomm used its Snapdragon Summit to show PrismML's 1-bit Bonsai LLM. The model runs locally on AI smart glasses built on the Snapdragon AR1 Gen 1 Platform. The processing happens on the glasses themselves and does not depend on a remote server.

This version of Bonsai has 2 billion parameters and is tuned for both vision and language. In practice, a wearer can ask about the scene in front of them and get an answer in real time. The model has to combine what the camera sees with what the user asks.

That combination is the hard part. A text-only assistant on a phone deals with one kind of input. A pair of glasses has to handle images and language together, on hardware small enough to sit on someone's face.

Shrinking without losing much: how compression is the core of the pitch

PrismML is known for taking larger models and making them much smaller while keeping almost all of their performance on standard benchmarks. For the glasses release, the reduction is fourfold.

The "1-bit" label points to how far the approach pushes precision down. Conventional models store their weights at higher precision, which uses more memory and more work per calculation. Low-bit designs give up some of that precision to fit into tighter hardware budgets. The challenge, which researchers who have tried ternarising a language model know well, is to cut the size without the output falling apart.

PrismML's claim is that it has managed this trade-off, keeping benchmark results close to the original. For a wearable device, where every megabyte and milliwatt counts, that claim decides whether a model is useful or simply too heavy to ship.

Open weights on local hardware: why PrismML is betting against the compute race

The glasses model is part of a bigger plan. PrismML wants to build open-weight AI that runs directly on devices and makes better use of the computing power already inside them.

The company presents this as an alternative to the current model of AI access. Today, many users send their data to proprietary labs and have to trust those labs' privacy promises. The same labs keep demanding more and more compute to serve their models. PrismML's argument is that capable models on the device reduce both dependencies.

The reasoning is simple. If the model sits on the glasses, the question about what the wearer is looking at does not have to leave the device. And if the model is small enough to run on existing chips, it does not need new data centre capacity for every new user. It is a similar logic to why some organisations are bringing open-weight models in-house: control over the model and the data moves closer to the people using it.

Local hardware brings its own limits, though. Raw compute is only one part of the picture, since memory bandwidth also shapes how quickly a model can respond. Smaller, lower-precision weights ease that pressure, which is one reason compression matters so much for devices like glasses.

A demonstration, not yet a product: what is still missing

A model built for Qualcomm's wearable chip platform moves PrismML closer to its stated goal. Having a major chipmaker present the model on its own stage also gives the startup visibility with the hardware makers who build on Snapdragon.

But there is a clear gap between a showcase and a product. So far, no smart glasses running PrismML's model have been announced. The technology has been demonstrated on the platform, but it is not yet available to buy.

That leaves the key questions open. Will device makers choose to ship a compact, open-weight model instead of connecting to a large cloud model? And will a 2-billion-parameter system hold up in everyday use, beyond benchmark scores?

The answers will show whether specialisation can compete with scale in wearables. For a device whose main job is to explain the world in front of the user, the relevant question is not how large the model is. It is whether the model can answer quickly and privately without leaving the frame.