Liquid AI releases LFM3 for specialized multimodal applications

Liquid AI has released LFM3, a family aimed at applications that need a model adapted to a particular task. Multimodal here means working with different kinds of input, such as images and speech. The company offers text, vision, audio, combined-input and decision variants, and says each can be fine-tuned on one GPU using the developer’s data.

The vision model’s documentation identifies document, chart and app-screen reading, plus questions about video, as intended tasks. It also explicitly discourages coding, multi-turn conversations and complex reasoning. Deployment support needs attention: llama.cpp integration is still under review and currently excludes video; SGLang support is under review, while vLLM support is in progress. The example commands therefore should not be read as confirmation that those released runtimes already support this model.

Omni’s instructions require both load_audio_from_video=True and use_audio_in_video=True to process a video’s soundtrack. Without them, only its frames reach the model.

In the tool example, the application parses and executes generated calls. It then appends both the assistant’s tool_calls record, with arguments as a dictionary, and a tool-role message containing results to the conversation. Only then does it generate again with the same tool definitions. The model can request more calls or answer; generating a call does not execute it.