Liquid releases LFM3 weights for AI that sees and hears on devices

Liquid AI released open weights for LFM3 on October 27. Six small models cover text, vision, audio, combined inputs and structured decisions, giving developers downloadable models for local inference on phones and laptops, with single-GPU customization.

The text model, LFM3-1.7B, has 1.73 billion parameters and supports 16 languages, including Korean. Liquid describes its four-bit memory footprint as about 1 GB. Its 256K-token context covers input and output together. Training targets local agents that call tools and return structured answers from documents.

Video input is new to the company’s vision-language range. LFM3-VL-2B can handle questions about recorded lectures or phone footage, read documents and charts, and locate objects. Its model card specifies 1.92 billion parameters, a 128K-token context and video sampling at two frames per second.

The audio model accepts speech or text and produces text, speech or both. It supports transcription, synthesis and translation across 18 languages, including Korean, with a 32K-token context. Streaming microphone processing and speech generation support spoken interaction. Spoken instructions can also be converted into tool calls for applications or devices. Audio input uses 16 kHz; generated speech uses 24 kHz.

LFM3-Omni-2B handles spoken questions about photographs and questions that need both a video’s images and its soundtrack. The 2.17-billion-parameter model combines jobs otherwise handled by separate speech-recognition, vision-language, language-generation and speech-synthesis models. It accepts text, images, video and audio, and answers in text or speech. Its 32K-token context is shorter than the text-only model’s maximum.

The two decision models return typed answers and probabilities in one forward pass, rather than generating a long response token by token. Their tasks include extracting numbers and dates with source spans, locating objects and estimating quantities. The audio-capable d1-Omni-2B targets calls and voice messages; its card warns that long spoken passages lose accuracy, especially outside Latin-script languages. For inputs without audio, Liquid identifies d1-2B as stronger on text and single bounding boxes.

Task limits and memory requirements

Liquid sets narrower expectations for the small text model: tool use, structured output and task-specific adaptation. Its card advises against agentic coding and complex multistep reasoning, and recommends retrieval or search for knowledge-heavy tasks. These limits accompany the release of the weights.

The announcement specifies 1.6 GB of KV cache for a full 128K-token context, adding memory beyond the roughly 1 GB model footprint. Its device speed comparisons use four-bit quantization and a 32K-token context.

Customization examples in the vision model’s card include extracting structured data from receipt photographs and converting images of equations into LaTeX. These workflows adapt the model to particular source material and output formats. The same card cautions against coding, multiturn conversations and complex reasoning.

Use is governed by the LFM Open License v1.0. Liquid says commercial users whose annual company revenue exceeds $10 million need a separate paid license. Qualified nonprofits using the models for noncommercial or research purposes are exempt from that threshold. Licensing downloaded weights is separate from paying for hosted API calls.

Fine-tuned models and modifications can remain private. Liquid retains the original model, while users own their contributions; the derivative remains subject to the license’s conditions. Redistribution requires a copy of the license and preservation of the applicable copyright and other notices.