Liquid AI has introduced a new vision-language foundation model known as LFM2-VL, which is tailored for efficient deployment across various hardware, including smartphones, laptops, wearables, and embedded systems. This model is designed to deliver low-latency performance, strong accuracy, and adaptability for real-world applications.
LFM2-VL is an iteration on the company’s previous LFM2 architecture, launched over a month ago, and claims to offer some of the fastest on-device foundation models available. It utilizes a linear input-varying (LIV) system that generates model settings dynamically for each input, facilitating multimodal processing for both text and images at different resolutions. According to Liquid AI, LFM2-VL can achieve up to twice the GPU inference speed compared to similar models while maintaining competitive performance benchmarks.
The release includes two model variants: LFM2-VL-450M, designed for resource-constrained environments with fewer parameters, and LFM2-VL-1.6B, which is lightweight enough for deployment on single-GPU systems. Both models can process images natively at resolutions up to 512×512 pixels and include features for handling larger images through a patching system that balances detail and contextual information.
Liquid AI is a company formed by former researchers from MIT’s Computer Science and Artificial Intelligence Laboratory, focusing on developing AI systems that move beyond traditional transformer models. Their flagship Liquid Foundation Models leverage principles from various scientific domains to create efficient, adaptable models that perform well with fewer computational resources.
The LFM2-VL models are now available on Hugging Face, including example fine-tuning code in Colab. They adhere to a custom “LFM1.0 license,” which allows for commercial use under specific conditions, intended to enhance accessibility for on-device and resource-limited applications.
Source: https://venturebeat.com/ai/liquid-ai-wants-to-give-smartphones-small-fast-ai-that-can-see-with-new-lfm2-vl-model/