Today, we release LFM2.5-VL-450M our most capable vision-language model for edge deployment. It processes a 512×512 image in 240ms and it is fast enough to reason about every frame in a 4 FPS video stream. It builds on LFM2-VL-450M with three new capabilities:

• bounding box prediction (81.28 on RefCOCO-M)
• multilingual visual understanding across 9 languages (MMMB: 54.29 → 68.09), and
• function calling support.

Most production vision systems are still multi-stage: a detector, a classifier, heuristic logic on top. This model does it in one pass:

• locating objects
• reasoning about context, and
• returning structured outputs directly on-device.

It runs on Jetson Orin, Samsung S25 Ultra, and AMD 395+ Max. Open-weight, available now on Hugging Face, LEAP, and our Playground.

HF model checkpoint: https://huggingface.co/LiquidAI/LFM2.5-VL-450M
Blog post: https://www.liquid.ai/blog/lfm2-5-vl-450m