LFM2.5-VL-3B Enhances Edge Vision

Liquid AI released LFM2.5-VL-3B, a 3.1B parameter vision-language model designed for on-device execution. It is designed to answer directly without generating lengthy reasoning traces, aiming for fast responses in real-time applications.

Key improvements include:

• Screen and UI understanding across various devices
• Object detection and grounding based on natural language queries
• Information comparison and understanding across multiple images
• Function calling supporting both text and image contexts

The model combines the SigLIP2 400M NaFlex vision encoder with the pre-trained backbone of the LFM2.5-2.6B text model. It was pre-trained on approximately 34 trillion tokens, with vision data increased fourfold compared to the previous version. The tokenizer vocabulary was also expanded to 128K to support non-Latin scripts.

Post-training involved knowledge distillation using a large teacher model, followed by SFT and Antidoom training, and finally multi-reward reinforcement learning. In official comparative evaluations, the average score across all vision tasks was 69.4, higher than the previous LFM2-VL-3B's 57.2.

Notably, it recorded 78.7 on ScreenSpot-v2 for desktop, 81.2 for mobile, and 82.2 for web, with a RefCOCO grounding score of 87.9. In tool-use evaluations, it scored 59.5 on ToolSandbox and 32.5 on BFCL V4, highlighting the potential for agent use in small models.