Liquid AI’s LFM3 brings six 2B-class models to edge devices

LFM3 spans text, vision, audio, omni and decision tasks with downloadable weights, but its $10 million revenue threshold makes the license less permissive than Apache 2.0.

2 min read

A compact single-board computer with visible ports and heatsinks, illustrating edge AI hardware.

Liquid AI released LFM3 on October 27 as a six-model family spanning base, text, vision, audio, combined audiovisual and calibrated decision tasks. The first checkpoints are in the roughly 2-billion-parameter class and are available as downloadable weights on Hugging Face, through Liquid AI’s API and in its Playground.

The text checkpoint, LFM3-1.7B, has 1.73 billion parameters and a 256,000-token combined input-and-output context window. Liquid says a 4-bit build uses about 1 GB of memory, while every family member can be fine-tuned on one GPU. Published runtime options include BF16, FP8, NVFP4, GGUF, MLX and ONNX builds across GPU, CPU and Apple Silicon workflows.

Liquid’s benchmark table reports 38.42 on WebArena-Verified for LFM3-1.7B, compared with 3.02 for its 2.6B-parameter LFM2.5 predecessor. The company also reports broader gains across tool use, agent tasks and multimodal evaluations. These are vendor-run results using specified harnesses and sampling settings; they do not establish universal superiority or independent replication. Liquid’s own model card cautions that the small text model is not recommended for agentic coding or complex multi-step reasoning.

A compact single-board computer represents local and edge deployment of small AI models.

Photo by Stefan Cosma on Unsplash, used under the [Unsplash License](https://unsplash.com/license/). Illustrative image, not an LFM3 device or benchmark result.

The weights use the LFM Open License 1.0, an Apache-2.0-derived license with a major restriction: free commercial rights end when a legal entity reaches $10 million in annual revenue, after which a separate commercial license is required. The release is therefore open-weight, but not standard permissive open source.

No canonical LFM3 source-code repository was identified. The official text-model page showed 211 Hugging Face likes and 6,311 downloads in the preceding month when checked October 28. Those are Hub totals, not GitHub stars or measured growth.

Sources: Liquid AI announcement · LFM3-1.7B model card · LFM Open License

Liquid AIOpen-weight modelsEdge AIMultimodalLFM3