6 models from one base
Liquid AI on Oct. 27 (Oct. 28 Japan time) released the weights of "LFM3," a family of small AI models that run on smartphones and computers. @aitrackerbot, an X account that tracks new model releases, was among the first to report their posting on Hugging Face.
All six models are built on the text model "LFM3-1.7B" (1.7 billion parameters). The text model fits in about 1GB when compressed to 4 bits and can take in about 260,000 tokens at once. "Omni," which handles audio, images, video and text together, is the company's first such model.
| Model | Input | Output |
|---|---|---|
| LFM3-1.7B | Text | Text |
| LFM3-VL-2B | Text, images, video | Text |
| LFM3-Audio-2B | Text, audio | Text, audio |
| LFM3-Omni-2B | Text, images, video, audio | Text, audio |
| LFM3-d1-2B | Text, images, video | Judgments with confidence scores |
| LFM3-d1-Omni-2B | Text, images, video, audio | Judgments with confidence scores |
Combining 3 types of layers
LFM3-1.7B's 24 layers consist of 13 convolution layers that look only at a short range, 5 attention layers that look only at the nearest 256 tokens and 6 attention layers that look at the whole input. Reducing the number of layers that look at the whole input keeps computation and memory use down even for long text. The company said it chose the design while actually measuring speed on 14 combinations, including smartphones, Raspberry Pi, computers and laptop processors.
The day before the release, the design and model names had already been spotted in a proposed change to llama.cpp, public software for running AI.
Claims to beat a 9-billion-parameter model
In Liquid AI's evaluation, LFM3-1.7B scored 49.37 on average across 29 benchmarks, above the 47.24 of "Qwen3.5-9B" (9 billion parameters) from China's Alibaba. It also pulled ahead on benchmarks of tasks using tools, the company said. However, the company does not recommend it for delegating programming or for complex multistep reasoning.
| Model | Average of 29 benchmarks |
|---|---|
| LFM3-1.7B | 49.37 |
| Qwen3.5-9B | 47.24 |
Commercial use by large companies requires payment
The license is the "LFM Open License v1.0," under which companies with annual revenue under $10 million can use the models commercially for free. Larger companies need a paid license from Liquid AI, and the same terms apply to modified models. This differs from open-source licenses that allow unrestricted use.
