Liquid AI announced LFM3 on Oct. 27, introducing six models for text, vision-language, audio, combined modalities and decision tasks. The company is pitching on-device use: It says users can fine-tune on one GPU, customize a model and run it on a device. It also describes an architecture optimized for 14 processor classes across GPUs, CPUs and NPUs.
The lineup starts with LFM3-1.7B, a text model that Liquid AI says has a 256k context window. LFM3-VL-2B is a vision-language model with video-input support, while LFM3-Audio-2B is described as supporting tool calls, voice chat and reasoning in 18 languages. The company says LFM3-Omni-2B combines vision, audio and language.
The remaining models, d1-2B and d1-Omni-2B, are for decision tasks. Liquid AI says they produce typed outputs and calibrated probabilities in a single pass, with the Omni version adding audio input. That is the company's description, not an independent finding that the probabilities are well calibrated.
In a separate post, @atomic_chat_hq said it had released four quantized versions of LFM3-1.7B on Hugging Face, ranging from 1.8GB to 1.1GB. It claimed all four fit on an 8GB Mac and recommended the 1.8GB Q8_0 version as the closest match to the original model. The hardware fit and comparison are the poster's claims. In a quote-post, @omarsar0 called the release unusually notable and said the models have “huge potential for always-on local agents.”
Comments