Liquid AI's LFM2.5-VL-3B: Fast On-Device Vision AI

Liquid AI has launched its new vision model, LFM2.5-VL-3B, bringing fast AI capabilities directly to devices. This model offers strong visual understanding for screens, user interfaces, and multi-image analysis. What's interesting is it runs efficiently, using only about 3 gigabytes of memory. For example, it achieves 228 tokens per second on an Apple M5 Max and 20 tokens per second on a Galaxy S26 Ultra. This new model improves significantly over its predecessor in areas like screen understanding and image grounding. It also supports various platforms, making it versatile for local AI assistants and visual agents. Meanwhile, Google DeepMind is enhancing accessibility with its SL2T model, translating over 50 sign languages into streaming text for smartphones. Microsoft also pushed its MAI-Image-2.6 model to second place on the Arena leaderboard for image generation. The bottom line is AI models are becoming smaller, more accessible, and more powerful across vision, communication, and creative tools.