Liquid AI Releases 6 Open-Weight Multimodal Models

The LFM3 family covers text, vision, audio, omni, and decision tasks. Each model can be fine-tuned on a single GPU and deployed on phones, laptops, private servers, or the cloud.

The LFM3 models are designed for efficient operation, with LFM3-1.7B requiring approximately 1 GB of memory at 4-bit precision. Checkpoints are available in GGUF, MLX, ONNX, FP8, and NVFP4 formats. The release includes specific models for local agents (LFM3-1.7B), vision with video input (LFM3-VL-2B), multilingual speech processing (LFM3-Audio-2B), and combined modalities (LFM3-Omni-2B).

Decision models, d1-2B and d1-Omni-2B, are also part of the release, capable of answering multiple questions in a single forward pass. The models feature a tri-hybrid architecture with 24 layers, incorporating gated short convolutions, sliding-window attention, and global attention layers. This architecture was developed through hardware-in-the-loop architecture search, building on lessons learned from LFM2.

Performance evaluations show LFM3-1.7B averaging 49.37 across 29 benchmarks, surpassing Qwen3.5-9B's score of 47.24 in Liquid's own evaluation. Notable performance metrics include 96.66 on IF Eval and 41.29 on AutomationBench, significantly outperforming Qwen3.5-9B's 15.10 on the latter. However, performance on Humanity’s Last Exam (without tools) was 7.17, and RULER 128K was 65.56, compared to Qwen3.5-9B's 87.11.