Liquid AI Releases LFM2.5-VL-3B for Faster On-Device Vision AI

Liquid AI has introduced LFM2.5-VL-3B, a 3.1-billion-parameter vision-language model designed to run on users’ own hardware. The community announcement describes it as the company’s most capable self-hosted vision model, with an emphasis on fast responses for edge and real-time applications.

The model adds stronger understanding of screens and user interfaces, natural-language object grounding, multi-image analysis, and function calling across both text-only and vision-plus-text workflows. Rather than producing extended reasoning traces, it is designed to answer directly—an approach intended to reduce latency in interactive applications.

Built for practical multimodal workflows

LFM2.5-VL-3B combines Google’s 400-million-parameter SigLIP2 NaFlex vision encoder with the pretrained backbone used by Liquid AI’s LFM2.5-2.6B text model. Its training included roughly 34 trillion tokens and four times more vision data than the company’s earlier releases, covering image captions, OCR, grounding, and instruction-following examples. Liquid AI also expanded the tokenizer to 128,000 entries to improve support for non-Latin scripts.

Post-training used supervised fine-tuning, knowledge distillation, Antidoom training, and a second phase based on multiple reward signals.

On the reported evaluations, the model led its size class on several real-world image tasks and showed competitive results in document, chart, OCR, multilingual, object-detection, and screen-understanding tests. Developers should still inspect the full benchmark table for task-specific trade-offs: results vary considerably by dataset, and some larger models score higher on individual evaluations.

The release is available through Hugging Face, giving builders another relatively compact option for privacy-sensitive multimodal agents, local document tools, UI automation, and edge deployments.

Source: Hugging Face Blog