Liquid AI launches Nanos for on-device tasks

Liquid AI announced Liquid Nanos, a family of compact foundation models with 350M to 2.6B parameters. The company says they reach frontier-model quality on specialized, agentic tasks while running entirely on phones, laptops and embedded devices. The models belong to its LFM2 family and use 100MB to 2GB of RAM.

The company argues that cost, energy, latency and privacy make it hard to send every request to the cloud. Its alternative is a small model fine-tuned for one job and running on the device. In evaluations with partners, Liquid says Nanos "perform competitively with models hundreds of times larger." The blog post does not name those partners.

The first models are labeled experimental:

• LFM2-Extract (350M and 1.2B): pulls data from unstructured text into JSON, XML or YAML, with inputs in nine languages
• LFM2-350M-ENJP-MT: translates between English and Japanese in both directions
• LFM2-1.2B-RAG: answers questions over long contexts in RAG pipelines
• LFM2-1.2B-Tool: handles function calling and agent tool use
• LFM2-350M-Math: solves math problems with step-by-step reasoning
• Luth-LFM2: a community-built series of French fine-tunes for on-device chat

All of the models are on Hugging Face and work with Liquid's LEAP and Apollo tools.

Liquid published details only for the extraction models. They were trained mostly on synthetic data and tested on 5,000 documents covering more than 100 topics. The scoring checked whether outputs parse correctly, match the requested format and stay faithful to the source text, and also used ratings from an LLM judge. Liquid says LFM2-1.2B-Extract beats Gemma 3 27B, a model 22.5 times its size, at producing complex structured outputs in several languages. It did not publish the scores, and the excerpt contains no benchmark results for the other Nanos.