Build specialized multimodal models with LFM3

Liquid AI has released LFM3, a new generation of open-weight models described in an October 27, 2026 post on Hugging Face. The family puts six models of roughly 2 billion parameters on one shared backbone, LFM3-1.7B-Base: a text model, the vision model LFM3-VL-2B, LFM3-Audio-2B, the company's first omni model LFM3-Omni-2B, and two decision models, d1-2B and d1-omni-2B. Each can be fine-tuned on a single GPU, which Liquid pitches as a route to small specialized models instead of one large generalist.

The backbone has 24 layers mixing 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers. It was pre-trained on 22 trillion tokens, and the text model offers a 256K context window and fits in about 1 GB at 4-bit precision.

Liquid's own benchmarks claim large gains. It says LFM3-1.7B beats LFM2.5-2.6B with 35% fewer parameters, scoring 38.42 against 3.02 on WebArena-Verified. The vision model reportedly lifts its average across 62 benchmarks from 57.6 to 71.4, which the company puts level with Gemma 4 12B. New in-house encoders include an audio design with time to first token as low as 50 ms.

The company also adapted the vision model for robotics, reporting 96.2% average success across the four LIBERO suites and action generation in under 50 ms on a Jetson AGX Orin. If these self-reported results hold up, capable agents and multimodal assistants could run on phones, laptops and embedded hardware rather than in the cloud.