LFM3: six 2B-class open models from Liquid AI that run on phones and a Raspberry Pi
What is it?
LFM3 puts six models on one LFM3-1.7B-Base backbone: text (LFM3-1.7B), vision (LFM3-VL-2B), audio (LFM3-Audio-2B), an omni model that takes text, images, video and audio (LFM3-Omni-2B), plus d1-2B and d1-omni-2B for fast structured decisions. It was pretrained on 22T tokens with a 256K context window, and the text model is about 1GB at 4-bit, small enough for phones or a Raspberry Pi 5. You can load the models straight from Transformers.
pip install -U transformersWhy does it matter?
Liquid AI says the text model scores 38.42 on WebArena-Verified, far above LFM2.5-2.6B's 3.02Vendor claim, and that the omni model beats larger models such as Qwen2.5-Omni-7B on some benchmarksVendor claim. If your app needs offline tool calling or structured output on the device, it belongs on the shortlist, and every model can be fine-tuned on a single GPU.
Maturity
GAWeights are on Hugging Face, and the models are also on the Liquid AI API and playground. The license is Liquid's own LFM Open License 1.0 rather than a standard one like Apache, so check the commercial terms yourself.
How to track it
- License: read the commercial-use terms, such as revenue thresholds, in the LFM Open License 1.0 on the model cards.
- Independent numbers: compare community llama.cpp and MLX ports and on-device speed tests with Liquid AI's figures.
- What's next: watch the Liquid AI blog and the Hugging Face collection for larger LFM3 models or new variants.