PrismML's 1-bit Bonsai model runs a VLM on smart glasses

At Qualcomm's Snapdragon Summit this week, PrismML demonstrated a 2-billion-parameter vision-language model running locally on smart glasses built around the Snapdragon AR1 Gen 1 platform. The model uses strictly binary weights — +1 and -1 — and executes on the Hexagon NPU. The headline is not the eyewear. It is that on-device inference keeps getting cheaper in the only unit that matters to an operator: memory per unit of capability.

What actually happened

Per PrismML's announcement and TechCrunch's report, the demo model is a 2B vision-language model built on PrismML's Bonsai 1.7B. At 1-bit precision it delivers what the company says is the conversational and visual reasoning of a comparable 4-bit model, while using roughly 4x less memory and generating tokens more than 2x faster.

PrismML is a Caltech-founded lab led by CEO Babak Hassibi and advised by UC Berkeley's Ion Stoica. Both companies framed the work as a privacy argument: "capable intelligence has to fit into the devices people already carry and wear," Hassibi said.

The part you can actually use today is the Bonsai repository, which ships Apache-2.0 weights at 1.7B, 4B, 8B, and 27B across three families — 1-bit Bonsai, ternary Bonsai, and Bonsai 2 — with backends for CUDA, Vulkan, ROCm, Metal, MLX, and plain CPU. Per the repo, Bonsai 2 27B packs to roughly 5.9 GB, keeps 262K context, and claims 98.2% of FP16 intelligence. The 27B variants take images.

Two things we are not going to claim. The glasses model itself was a demo, not a shipping product, and no eyewear running it has been announced. And the 98.2% figure is vendor-reported — run it on your own evals before you believe it.

Why on-device AI matters for your business

A 27B-class model in 5.9 GB changes what hardware you need. That is a mini-PC in a back office, not a GPU lease. Every classification, extraction, and routing call you currently send to a hosted API is a line item that grows with your volume. The same call against local weights has a marginal cost of zero and a latency floor set by your own silicon.

Apache-2.0 is the clause that survives a pricing change. We have written about retired model aliases and endpoints that simply vanished often enough to be blunt: a permissively licensed file on your own disk cannot be deprecated, throttled, or repriced mid-quarter. Even as a fallback tier behind a frontier model, it is cheap insurance.

Aggressive quantization is a routing decision, not a religion. One-bit weights buy memory and speed, and they cost accuracy somewhere. Find where. Pick two or three tasks you already run at volume, build a 50-example eval set from your real data, and measure the local model against your hosted one. If it holds, route those tasks local and keep the expensive model for the work that earns it.

Sources: PrismML announcement, TechCrunch, Bonsai on GitHub.