The new LFM3 family uses a shared architecture that runs locally on devices, cutting latency and addressing privacy while handling text, image and audio tasks.

Liquid AI has announced the LFM3 suite, a family of multimodal models that can run entirely on‑device, handling text, image and audio inputs without needing cloud connectivity.

Shared architecture for on‑device AI

The LFM3 models use a unified architecture that processes different data types through the same core network, allowing developers to deploy a single model for multiple tasks on smartphones, tablets and other edge devices.

Performance and privacy benefits

Running locally eliminates round‑trip latency to remote servers, delivering faster responses for interactive applications. Because data never leaves the device, user privacy is better protected, a key selling point for health, finance and personal assistant use cases.

Supported modalities

  • Text generation and understanding
  • Image captioning, classification and editing
  • Audio transcription, synthesis and classification

Developers can combine these capabilities in a single app, for example by generating a spoken summary of an image caption or by providing real‑time transcription of a video feed.

Liquid AI says the suite is optimized for a range of hardware, from flagship smartphones to low‑power IoT devices, using quantization and pruning techniques to keep memory footprints modest.

For readers interested in broader AI developments, the AI category on uxc.news offers daily coverage of breakthroughs and industry trends.

The ability to run multimodal AI locally is a game‑changer for user experience and data security.

The announcement was detailed in Liquid AI’s press release and covered by AI Adapted’s October 29, 2026 article.

Track coverage on Momentum · More in AI.