Liquid AI releases LFM3, a small open-weight multimodal model family

Liquid AI has released LFM3, a new generation of its Liquid Foundation Models built for running on-device and being fine-tuned for specialized tasks. The first release covers six models around 2 billion parameters each, spanning text, vision, audio, and omni inputs, plus two decision models, all branched from a shared LFM3-1.7B-Base backbone trained on 22 trillion tokens.

The family uses a new tri-hybrid architecture that mixes gated short convolutions, sliding-window attention, and global attention across its layers, tuned using hardware-in-the-loop search across 14 device and backend combinations, including phones, a Raspberry Pi, laptops, and cloud chips. The text model, LFM3-1.7B, has a 256K-token context window and fits in about 1 GB at 4-bit precision.

Liquid AI's own benchmarks claim LFM3-1.7B outperforms its previous LFM2.5-2.6B model while using 35% fewer parameters, and that it beats much larger models like Qwen3.5-9B on some agentic browser and business-app tasks, though larger models still lead on coding and general knowledge. The vision model, LFM3-VL-2B, is claimed to match Gemma 4 12B, a model over six times its size, on average across 62 vision benchmarks. All models and weights are available now on Hugging Face and the Liquid AI API, with partners including AMD, Ambarella, and Qualcomm highlighting on-device use cases.

Why it matters

Small, open, on-device models that can be fine-tuned on a single GPU let developers build specialized AI features without relying on constant cloud calls.

Source: Liquid AI