LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge

Details

Liquid AI's LFM2.5-VL-3B pairs a SigLIP2 400M NaFlex vision encoder with the pre-trained LFM2.5-2.6B text backbone, pre-trained on approximately 34 trillion tokens — 4x more vision data than prior Liquid AI releases. Post-training runs in two stages: supervised fine-tuning with knowledge distillation from a larger teacher model and Antidoom training, followed by multi-reward reinforcement learning.\n\nFour capability areas are highlighted: screen/UI understanding across device types, improved grounding with natural language queries for object detection, multi-image input reasoning, and function calling in both text-only and vision-text scenarios. The model answers directly rather than reasoning through chain-of-thought steps, keeping response latency low for real-time and on-device applications.\n\nBenchmark results show LFM2.5-VL-3B leading its size class on real-world image tasks: 63.3 on MMStar, 73.1 on RealWorldQA, 68.5 on MathVista (mini), 60.6 on MM-IFEval. It outperforms Gemma 4-E4B-it (8B params) on MMStar, RealWorldQA, and other tasks while using fewer than half the parameters. The vocabulary was doubled to 128K tokens for non-Latin script support by extending the tokenizer in place, without retraining from scratch. The model is available on Hugging Face for local deployment.