LFM3

LFM3 is Liquid AI's open-weight family of small models for text, vision, audio, and decisions, for people who want to run one on their own device.

Six checkpoints from one base

The announcement is dated October 27, 2026. It says the first release sits around 2 billion parameters and branches from LFM3-1.7B-Base. Weights are on Hugging Face, there is a playground, and there is a Liquid API console. This page did not download a checkpoint or open that console, so API prices are not here. The weights themselves are what the post calls open-weight.

• LFM3-1.7B is the text model. The card lists 1.73 billion parameters and a 256K context. It says 4-bit weights take about 1 GB. The languages on that card are English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, and Polish. The license line is LFM Open License v1.0.
• LFM3-VL-2B reads images. The post says its vision encoder uses 3 times less compute than the SigLIP2 encoder in their previous vision models.
• LFM3-Audio-2B takes text or audio and answers with text, speech, or both. The post says it supports 18 languages.
• LFM3-Omni-2B is described as the first model in the family that can see, listen, and talk.
• d1-2B is a decision model built on the vision model. d1-Omni-2B also listens. The post says they do not write an answer. They score a list of questions in one pass, each with a probability.

The image on this page is the announcement's own family graphic, not a generation from here. It labels five of them: LFM3-1.7B as text, LFM3-VL-2B as vision, LFM3-Audio-2B as audio, LFM3-Omni-2B as omni generative, and d1-Omni-2B as omni decision. d1-2B is in the post and is not on that graphic.

The card draws a line the tables do not

The text model card says the checkpoint is for local agents, tool calling, structured output, and a short fine-tune on one GPU. It says to use retrieval for knowledge-heavy tasks, and it is not recommended for agentic coding or complex multi-step reasoning because of its size. The same announcement claims LFM3-1.7B outperforms their previous LFM2.5-2.6B while using 35 percent fewer parameters, and it prints comparison tables. Those scores are the post's.

The architecture paragraph says 24 layers: 13 gated short convolutions, 5 sliding-window attention layers with a 256-token window, and 6 global attention layers. Pre-training is given as 22 trillion tokens. The text model is described as post-trained for harnesses named Hermes Agent, Pi, and OpenClaw.

One fine-tune they report

A demo in the same post fine-tunes LFM3-VL-2B on about 3,000 public wind-turbine photos, for a few hours on one A100. They say the untouched model and GPT-6 Luna identify those defect types in fewer than 10 percent of about 1,400 held-out photos, and the fine-tune reaches 67 percent classification accuracy, with detection F1 moving from 0.003 to 0.11. Those figures are theirs. The notebook is in the Liquid4All cookbook. Weights for the family are in the Hugging Face collection. The accounts linked from the site are the GitHub org and @liquidai.