A single backbone for five model families
Liquid AI has presented LFM3, the new generation of its Liquid Foundation Models, described on the blog hosted by Hugging Face as the basis for what the company calls specialized intelligence. The central element is a backbone shared by text, vision, audio, omni and decision models, all on the order of 2 billion parameters. According to the source, this makes it possible to choose the model suited to the input modality, adapt it with fine-tuning on a single GPU, and deploy it where the application runs.
The novelty compared with LFM2 is not only the addition of modalities: Liquid AI claims to have redesigned the hybrid architecture, which combines short convolutions and attention, through hardware-in-the-loop research for agentic use cases. The research now covers fourteen device and backend combinations, from phones to Raspberry Pi 5, laptops, cloud CPUs, GPUs and Qualcomm Hexagon NPUs. This is a company statement, not independently verified data.
Text and vision: the claimed figures
The flagship text model is LFM3-1.7B: according to the source, it is post-trained to work in agentic environments and would outperform the predecessor LFM2.5-2.6B with 35% fewer parameters, with the largest gains on agentic tasks. On WebArena-Verified, LFM3-1.7B would score 38.42 versus 3.02 for the predecessor, and would beat Qwen3.5-9B on long browser tasks and enterprise applications. These are results reported by Liquid AI in its own technical data sheets.
On the vision side, LFM3-VL-2B is the company's first vision model that accepts video as input and reasons step by step before answering. With 1.9 billion parameters, versus 3.1 for the predecessor, the average across 62 vision benchmarks would rise from 57.6 to 71.4. The source describes it as a first step toward robotics, with three derived models for orchestration, navigation and manipulation; however, the results on LIBERO and RoboCasa-GR1 are presented by the company.
Audio, omni and decision-making
LFM3-Audio-2B focuses on voice conversation, recognition, synthesis and translation in 18 languages on laptops and edge devices. Compared with LFM2.5-Audio-1.5B, it introduces an internally developed audio encoder and voice codec, with claimed higher quality (UTMOSv2 3.83 versus 3.22) and new capabilities for audio understanding and tool calling. LFM3-Omni-2B, meanwhile, is the company's first omni model, with text, vision and audio input and text and voice output, through the LiquidCodec codec, which would generate speech faster than real time.
The d1-2B and d1-omni-2B family is designed for structured decisions: it receives a situation and a list of questions and answers in a single pass, with calibrated probabilities. The models are open-weight and available on Hugging Face. The actual equivalence of the claimed performance relative to the cited models remains to be independently verified, especially because the comparisons come from the same company's documentation.
