Liquid AI Releases LFM2.5-VL-3B, a 3.1B VLM That Reads Screens and UIs - Fits in About 3GB

Liquid AI released the vision-language model LFM2.5-VL-3B on August 12, 2026¹. It has 3.1 billion parameters, pairing a SigLIP2 400M NaFlex vision encoder with the same backbone as the company’s text model¹. Liquid AI lists four improvements: screen and UI understanding, grounding and object detection from natural language queries, multi-image input, and function calling¹.

The runtime numbers are concrete. It fits in about 3GB of memory and decodes 228 tokens/s on an M5 Max, 116 tokens/s on a Ryzen AI Max+ 395, and 20 tokens/s on a Galaxy S26 Ultra¹. It ships with day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX¹. The license is not a standard one such as Apache 2.0 but a custom “lfm1.0”². As Hugging Face’s analysis showing that sub-1B models account for most downloads suggests, this adds one more small on-device option — though the terms for commercial use require reading the license itself.

Sources

1. LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge - Liquid AI (Hugging Face blog, August 12, 2026)
2. LiquidAI/LFM2.5-VL-3B - Hugging Face model repository (license checked August 17, 2026)