Liquid AI releases LFM3: six open-weight ~2B models for on-device agents

On 27 October 2026, Liquid AI published LFM3, a family of open-weight models it describes as the starting point for specialised, on-device AI. According to the company's post on the Hugging Face blog, the family has six models of about 2B parameters that cover text, vision, audio and a combined omni model, plus two decision models, all trained from one shared 1.7B base.

What Liquid AI says is in the family

• LFM3-1.7B (text): aimed at local agents, tool calling and structured output. Liquid AI says it fits in about 1 GB at 4-bit precision and has a 256K-token context window.
• LFM3-VL-2B (vision): adds video input. Liquid AI reports an average of 71.4 across 62 vision benchmarks, up from 57.6 for LFM2.5-VL-3B, and says that is on par with Gemma 4 12B.
• LFM3-Audio-2B: voice chat, speech recognition, speech synthesis and translation across 18 languages. The company says time to first token can be as low as 50 ms.
• LFM3-Omni-2B: takes text, images, video and audio and answers in text or speech, intended to replace a chain of separate speech, vision, language and voice models.
• d1-2B and d1-omni-2B: decision models that return calibrated answers.

The agent numbers, and who ran them

Liquid AI reports that LFM3-1.7B scores 38.42 on WebArena-Verified, a test of browser and business-app tasks, against 3.02 for its predecessor LFM2.5-2.6B, and that it beats Qwen3.5-9B on long tasks in browsers and business apps. These are the vendor's measurements; the post does not say they were reproduced by anyone else. The company credits a new layer mix, a larger share of agent traces and tool calls in the final 3.5T tokens of training, and a 22T-token pre-training run.

Fine-tuning is the pitch

Liquid AI argues that frontier models are generalists and that a short fine-tune of a small model can make a reliable specialist. It says each model can be fine-tuned on a single GPU and links ten notebooks. Tasks named in the post include extracting structured JSON from messy documents, answering from retrieved documents or declining, calling the right tool in a 108-tool phone app, reading receipt photos into JSON, and mapping spoken smart-home commands to function calls.

For a team, that makes LFM3 a candidate for narrow, high-volume jobs where data should stay on a device or in a private network. The post does not report cost per task, accuracy on those fine-tuning tasks, or how far the benchmark gains hold on a company's own documents.

What the post does not say

• The licence terms and any commercial-use limits are not stated in the post.
• No independent evaluation is cited; every figure comes from Liquid AI.
• Some models need recent library versions: the post lists transformers 5.21.0 or later for the vision model and 5.20.2 or later for audio and omni, and the examples set trust_remote_code to true, which runs code shipped with the model.