Liquid AI releases LFM3 models
On 27 October 2026 Liquid AI released LFM3, a new generation of its Liquid Foundation Models: six models of roughly two billion parameters built on one 1.7-billion-parameter backbone, for text (LFM3-1.7B), vision with video input (LFM3-VL-2B), speech (LFM3-Audio-2B, with a new audio tokenizer, LiquidCodec), text, images, video and audio together (LFM3-Omni-2B) and two models that answer with calibrated probabilities (d1-2B and d1-omni-2B). The weights are on Hugging Face under the LFM Open License v1.0.
Why it matters
Small models for phones and laptops now come as one family that takes text, images, video and audio in and can speak back, with weights a developer can download. The licence is not open source in the usual sense: it permits commercial use only for companies below 10 million US dollars in annual revenue, a limit the announcement post does not name.
What the blog says. Blog post on Hugging Face of 27 October 2026 (a team article by Liquid AI). The text model LFM3-1.7B is trained from LFM3-1.7B-Base, a 24-layer mix of gated short convolutions, sliding-window attention and global attention, with a 256K-token context window, and "fits in about 1 GB at 4-bit precision". LFM3-VL-2B is "the first LFM vision model that takes video as input". LFM3-Audio-2B takes text or audio and returns text, speech or both, for voice chat, recognition, synthesis and translation "across 18 languages" on a laptop or edge device; its audio encoder is fully causal, so the model can prefill while the user speaks ("time to first token as low as 50 ms"), and LiquidCodec decodes speech "at about 87x real time on a single CPU thread on Apple M4 Max". LFM3-Omni-2B is the first omni-modal LFM. The two d1 models answer a list of questions in one pass, each with a calibrated probability. These are Liquid AI's own measurements and descriptions. The licence. The model cards carry the tag lfm1.0 and the LICENSE file is "LFM Open License v1.0": an Apache-style grant "subject to ... the Commercial Use limitation", under which "Any Commercial Use of the Work or a Derivative Work by a Legal Entity that exceeds the Threshold is not licensed"; the Threshold is annual revenue of 10 million US dollars or more. It does not apply to a qualified non-profit using the work for non-commercial or research purposes. The blog text does not name the licence. Dates. Each model repository's history begins on 27 October 2026; the repositories themselves were created between 8 September and 26 October. The base model LFM3-1.7B-Base is also published. What the record does not claim. The blog compares the models with earlier LFM models and with larger third-party models; those comparisons are Liquid AI's own and none is repeated as a finding.