Liquid AI releases LFM3 models built from one multimodal backbone
Liquid AI released the open-weight LFM3 family, bringing text, vision, audio and structured decisions onto a shared foundation. Developers can download models and adaptation notebooks for specialized tasks such as document extraction and spoken tool use. The release includes a common base for further training; its availability should be distinguished from opening the entire training process and data.
Artificial Intelligence··Midday
One foundation branches into different forms of AI
Liquid AI, a developer of compact AI models, released LFM3 with open weights for developers to download and adapt.[1], [2]
The family brings language, vision, audio and structured-decision models onto a common foundation. Rather than training every modality from an unrelated starting point, the models branch from LFM3-1.7B-Base. The release includes the base checkpoint as well as models adapted after initial training.[1]
The backbone combines three kinds of sequence-processing layer. Its 24 layers comprise 13 gated short convolutions, five sliding-window attention layers and six global attention layers. The short-convolution kernel grows from three in the earlier family to seven. Liquid AI’s architecture search covers 14 device and backend combinations, ranging from phones and Raspberry Pi 5 computers to laptops and cloud processors.[1]
Speech and vision connect to practical tasks
The vision variant, LFM3-VL-2B, accepts video input, a first for the company’s LFM vision models. The audio model accepts speech or text and can return speech, text or both. One demonstrated task turns a spoken request into a tool call. The combined LFM3-Omni-2B model processes text, vision and audio together, with text or speech as output.[1]
The d1 decision variants answer several questions in one forward pass and attach probabilities to structured answers. Their examples include choosing multiple applicable options, extracting document values and locating objects. This output format is aimed at explicit decisions and fields rather than an extended passage of generated prose.[1]
Adaptation notebooks accompany the downloadable weights
Liquid AI says each model can be fine-tuned on a single graphics processor. Published notebooks cover data preparation, adaptation and evaluation, with examples involving document extraction, learning Turkish, spoken tool use and receipt reading. Separate robotics examples illustrate navigation, coordination and action generation. These are demonstrations of adaptation paths. Downloadable weights do not by themselves open the complete training data or process, and the company’s internal performance comparisons are not independent measurements.[1]
