Meta releases Llama-4o-mini: sub-8B model with improved on-device layout understanding

Tech · 3 min read

Meta's Llama-4o-mini is a trimmed variant of its Llama-4o family, engineered for tasks like layout parsing, OCR, and UI element recognition on mobile devices. The model delivers strong performance on extracting component hierarchies from screenshots while keeping inference latency under 200ms on flagship phones.

Developers can use the model with Meta's new mobile SDK, which includes prebuilt adapters for popular design tools and a small footprint storage format for offline use. The SDK also offers privacy-preserving features, including on-device differential privacy for usage telemetry.

Designers and product teams anticipate using the model to power offline prototyping, automated spec generation, and quick usability checks during fieldwork. Privacy advocates welcome the local-first approach, although some experts note the tradeoffs in generalization versus larger cloud-hosted models.