Liquid AI releases LFM2.5-VL-3B, a 3-billion-parameter vision-language model tuned for on-device use

Liquid AI has released LFM2.5-VL-3B, a 3.1-billion-parameter open-weight vision-language model designed to run on user hardware. The company describes it as its most capable model to date at that size, with four main improvements over the previous LFM2-VL-3B release: stronger understanding of digital screens and user interfaces across devices, improved object detection and grounding from natural-language queries, better reasoning across multiple images, and much stronger function calling in both text-only and vision-text settings.

The model pairs a SigLIP2 400M NaFlex vision encoder with the same pre-trained backbone as the LFM2.5-2.6B text model. It was pre-trained on about 34 trillion tokens, with four times as much vision data as the previous version, drawn from curated and synthetic image-caption, OCR, grounding and instruction-following sets. To support non-Latin scripts, the vocabulary was doubled to 128,000 by extending the tokenizer in place rather than retraining from scratch. Post-training runs in two stages — supervised fine-tuning with distillation from a larger teacher, followed by multi-reward reinforcement learning.

On the ScreenSpot-v2 benchmark for reading on-screen elements, LFM2.5-VL-3B scored 78.7 on desktop, 81.2 on mobile and 82.2 on web tasks, ahead of most 2B and some 4B class models in Liquid AI's evaluation. On RefCOCO grounding it scored 87.9, up from 57.1 for the previous version. Instruction-following and tool-use scores also improved sharply, though it still trails some larger competitors on general instruction-following.

For inference speed, Liquid AI reports the model decodes 228 tokens per second on an M5 Max, 116 tokens per second on a Ryzen AI Max+ 395, and reaches 20 tokens per second on a Galaxy S26 Ultra so it can run fully on-device. It uses about 3 GB of memory. On GPU, the company says it is the fastest of the models tested on multi-frame inputs and reaches about 11,000 tokens per second at high concurrency on an H100.

The model ships with day-one support in llama.cpp, MLX, vLLM, SGLang, ONNX and the latest Transformers library. It is designed to answer directly rather than through a chain of reasoning, so responses stay fast in real-time and on-device applications.