PrismML launches tiny LLMs for Qualcomm-powered smart glasses, enhancing offline AI

PrismML announced on Tuesday that its tiny language models are now running on Qualcomm‑powered smart glasses, allowing AI tasks to be processed directly on the device without a cloud connection. The integration works on the Snapdragon XR2 platform that powers most commercial AR headsets. Users will be able to ask the glasses for translations, captions or contextual information and receive answers instantly, even when offline.

Running AI on the edge matters because it cuts latency, saves bandwidth and protects privacy. As AR glasses move from niche prototypes to consumer products, manufacturers need to make the most of limited battery life and on‑board compute. Qualcomm’s XR2 chip offers a mix of CPU, GPU and AI accelerator cores, but developers have struggled to fit large language models within those constraints. PrismML’s approach is to shrink the models to a few hundred megabytes while keeping core language abilities, an effort that aligns with the broader industry push for “open‑weight” AI that can be customized and deployed on any hardware. The move also follows recent releases from Apple and Meta that showcase on‑device AI for vision and speech, signalling a shift away from always‑online services.

PrismML’s CEO, Anika Patel, said the partnership with Qualcomm began after the company released a 350‑million‑parameter model that could run on a laptop with under 8 GB of RAM. Qualcomm’s engineering team then optimized the model for the XR2’s AI engine, cutting inference time to under 200 ms per query. Analysts at IDC noted that this could accelerate the adoption of AR glasses in enterprise settings such as logistics, field service and remote assistance, where reliable, low‑latency AI is critical. PrismML plans to release a developer SDK later this month, enabling third‑party apps to embed its models. The company also hinted at a next‑generation model that will support multilingual understanding and more complex reasoning, aiming for a broader launch in early 2027. If successful, the technology could set a new standard for on‑device intelligence across the growing market of wearable computers.