Liquid AI Open-Sources Lightweight Multimodal Models: Extracting Images into JSON Structured Data Directly on Device

According to monitoring by Beating, Liquid AI has open-sourced two lightweight multimodal models, LFM2.5-VL-1.6B-Extract and LFM2.5-VL-450M-Extract. These new models are specifically optimized for extracting structured data from images, enabling direct conversion of images into JSON-formatted data on devices based on user-specified field lists. This eliminates the traditional step of generating full text with multimodal models and then performing secondary parsing. The new models are available in two parameter configurations: 1.6 billion (1.6B) and 450 million (450M), and are released under the LFM Open License v1.0. Official evaluations show that the new models perform exceptionally well in scenarios such as document scanning, in-cabin vehicle understanding, and industrial inspection. In benchmark tests, the 1.6B model delivers performance comparable to general-purpose multimodal models with 4 billion (4B) parameters, while the 450M model rivals models with 2 billion (2B) parameters. On the deployment side, the new models have been adapted for various smart hardware and edge device chips (SoC), enabling offline deployment in edge-side scenarios such as in-cabin vehicle understanding, document scanning, and industrial inspection. Liquid AI has made the model weights available for download on the Hugging Face platform. [Original Link]