Signals/ Brief /
Liquid AI's LFM3 puts six small models on phones and laptops
The models are about 2B parameters each, cover text, images, video and speech, and run locally, but they are not built for complex reasoning.
On October 27, Liquid AI released LFM3, a family of six open-weight models, each around 2 billion parameters. They are small enough to run on a phone or an ordinary laptop, with no cloud account needed.
The lineup covers text (LFM3-1.7B), images and video (LFM3-VL-2B), speech (LFM3-Audio-2B), and all of those at once (LFM3-Omni-2B). Two “decision” models return calibrated answers instead of free-form text. The text model handles a 256K-token context and 16 languages. The model card says it fits in roughly 1 GB at 4-bit precision.
What’s new:
- Weights are on Hugging Face, with a hosted playground and API as well.
- Launch partners include Arm, AMD, Intel, NVIDIA and Qualcomm.
The catch: Speed and quality numbers come from Liquid AI, and independent testing is thin. The model card does not recommend these for agentic coding or complex multi-step reasoning. Licensing is also limited: one guide says the text model’s license makes commercial use free only for companies under $10 million in annual revenue, and another source confirms the license only for that model. Read each model card before building a product on it.
How to try it
- Open the LFM3 page on Hugging Face and read the license on the model you want.
- Use a local runner. The hardware guide says recent llama.cpp builds work, while some older Ollama versions fail to load it.
- Start with the 4-bit text version, which is the smallest.
- Test it on one narrow job, like sorting emails or pulling fields from forms.
Should you switch?
- Yes, if you want private, offline helpers for simple, repeatable tasks.
- No, if you need deep reasoning or coding help. Use a larger hosted model.
