AI for your phone: Liquid AI releases six small open models in its LFM3 family
Text, images, video, speech and decisions: the US start-up's new models share a common backbone and are designed to run directly on smartphones and laptops.
Graphic: TERRA WORLD NEWS (own graphic based on the sources listed)
Video: YouTube · Liquid AI · youtube.com
US start-up Liquid AI on Tuesday unveiled LFM3, the new generation of its compact AI models. The family comprises six models in the two-billion-parameter class: a text model, a model for images and video, an audio model that can listen and speak, an "omni" model for every type of input and two so-called decision models that return calibrated answers instead of free text. All are derived from the same base model. Liquid AI's video introduces the lineup.
According to the company, the LFM3-1.7B text model handles up to 256,000 tokens of context, supports 16 languages and needs only about one gigabyte of memory in compressed form; the audio model supports 18 languages. The weights are openly available on Hugging Face, and Liquid AI says each model can be fine-tuned on a single GPU. Launch partners named by the company include AMD, Arm, Intel, Nvidia and Qualcomm.
The performance figures come from the developer itself: the omni model is said to beat larger models with up to 10.7 billion parameters in several tests. Independent comparisons are still pending. Liquid AI explicitly does not recommend the text model for agentic coding or complex multi-step reasoning tasks.