Liquid AI releases LFM3, small open-weight on-device models

Liquid AI on Oct. 27 released LFM3, a new family of small AI models for text, vision, audio and decision-making that it says are built to run on phones, laptops and other devices. The company said all six models branch from one shared base model, sit in the 2 billion-parameter range, and can each be fine-tuned on a single GPU. The weights are on Hugging Face now.

The launch post lists dozens of Liquid AI authors, among them Ramin Hasani, Mathias Lechner and Alexander Amini. Liquid AI says on its company page that it was spun out of MIT’s Computer Science and Artificial Intelligence Laboratory, or CSAIL, by four founders: Hasani, its chief executive; Lechner, its chief technology officer; Amini, its chief scientific officer; and Daniela Rus.

What’s in the family: Every model starts from LFM3-1.7B-Base, Liquid said. LFM3-1.7B handles text, with a 256,000-token context window, and is aimed at local agents, tool calling and structured output. LFM3-VL-2B adds images and is the first Liquid vision model to take video. LFM3-Audio-2B handles voice chat, speech recognition, speech synthesis and translation across 18 languages. LFM3-Omni-2B, the company’s first omni model, takes text, images, video and audio and answers in text or speech. Two “decision” models, d1-2B and d1-omni-2B, answer a list of questions about an input in one pass, each with a calibrated probability, the post says.

Built for devices: Liquid said the text model fits in about 1 GB at 4-bit precision. It described a new “tri-hybrid” design that mixes short convolutions with two kinds of attention, found through a search tested on 14 device and software combinations, from phones and a Raspberry Pi 5 to laptops, GPUs and the Qualcomm Hexagon NPU. With 8,000 tokens in context, LFM3-1.7B runs about 25% faster than LFM2-1.2B per parameter, the company said. Its new audio encoder can start answering in as little as 50 milliseconds, according to the post.

Liquid’s claims: The company said LFM3-1.7B outperforms its earlier LFM2.5-2.6B with 35% fewer parameters, scoring 38.42 against 3.02 on the WebArena-Verified agent test. It said LFM3-VL-2B, with 1.9 billion parameters against 3.1 billion for the earlier LFM2.5-VL-3B, raised the average across 62 vision benchmarks from 57.6 to 71.4, on par with Gemma 4 12B, a model more than six times larger. All of these results are reported by Liquid.

The license: The model cards list the LFM Open License v1.0, not an Open Source Initiative-approved license. It allows free use, modification and redistribution, but commercial use is licensed only for organizations with less than $10 million in annual revenue, according to the license text. Qualifying nonprofits are exempt for noncommercial and research use.

Why it matters: LFM3 is a bet on many small specialists instead of one big generalist. Liquid’s pitch is that a developer picks the model for the input, fine-tunes it on one GPU and ships it wherever the application runs, from devices to private servers, the post says, rather than routing every request through one large general-purpose model. The reported gains are large for the size, but they are Liquid’s own numbers, and outside testing will show how well they hold up. The license matters as much as the benchmarks: “open-weight” here means anyone can download and modify the models, but commercial use by an organization with $10 million or more in annual revenue isn’t licensed under these terms. That puts LFM3 within reach of hobbyists, researchers and smaller startups, while the license doesn’t cover commercial use by bigger companies.

What’s next: Liquid said the full LFM3 collection, including the base model for custom post-training, is available on Hugging Face today, along with fine-tuning notebooks and a guide to running the models in local agent harnesses such as OpenClaw, Hermes Agent and Pi.