Meta open-sources Palimpsest, a compact model for on-device AR experiences

Tech · 4 min read

Palimpsest focuses on spatial awareness, fast visual grounding, and low-bandwidth prompts, enabling AR apps to generate contextual UI suggestions without constant cloud calls. The model is quantized and engineered for mobile GPUs and dedicated NPUs, with a small memory footprint suited for consumer headsets and phones.

Alongside the model, Meta provided an SDK with example pipelines for label placement, dynamic callouts, and content occlusion management. The tooling includes developer controls for energy usage and latency trade-offs so apps can balance responsiveness against battery life.

Meta frames Palimpsest as part of a broader strategy to enable privacy-preserving AR interactions where most inference happens locally. The open-source license and sample apps aim to accelerate adoption across diverse AR ecosystems.