Liquid AI adds grounding to 450M vision model

Liquid AI announced LFM2.5-VL-450M, an updated version of its LFM2-VL-450M vision-language model. It extends the LFM2.5 family, whose on-device models we covered earlier. The new model adds grounding, better instruction following and function calling support.

The company scaled pre-training from 10T to 28T tokens. It then used preference optimization and reinforcement learning in post-training to improve grounding and reliability. The model now predicts bounding boxes, scoring 81.28 on RefCOCO-M; its predecessor had no detection capability. Liquid reports that multilingual image understanding on MMMB rose from 54.29 to 68.09 across eight languages, and MM-IFEval instruction following rose from 32.93 to 45.00. CountBench climbed from 47.64 to 73.31. Not every score went up: MMMU and InfoVQA dipped slightly against the earlier model. On text-only function calling, the model scores 21.08 on BFCLv4. Liquid's comparison table also includes SmolVLM2-500M, which trails on most benchmarks.

For latency, Liquid quantized the model to Q4_0. It processes a 512×512 image in 242ms on a Jetson Orin, which the company says is enough to handle every frame of a 4 FPS stream. The same image takes 944ms on AMD's Ryzen AI Max+ 395 and 2.4s on a Samsung S25 Ultra. At 256×256, the phone time drops to 950ms.

Liquid is targeting the model at industrial automation, wearables and high-throughput retail image processing. It is available on Hugging Face, LEAP and Liquid's Playground. Licensing terms were not stated in the announcement.