LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge

LiquidAI Releases 3B Vision Model for On-Device AI

LiquidAI has released LFM2.5-VL-3B, a 3.1 billion parameter vision-language model designed to run efficiently on local hardware. The model targets the growing demand for on-device AI by offering real-time capabilities for applications that process visual information from documents, screens, and multiple images without relying on cloud inference. This release positions LiquidAI as a direct competitor in the small, efficient model space, focusing on practical performance over sheer size.

Technical Improvements and Benchmark Standing

The new model integrates a SigLIP2 400M NaFlex vision encoder with the text backbone from the company's LFM2.5-2.6B model and was pre-trained on 34 trillion tokens. According to the company's documentation, its development focused on four key areas of improvement:

• Screen and UI Understanding: Enhanced comprehension of digital screens across devices.
• Improved Grounding: Better object detection and localization via natural language queries.
• Multi-image Reasoning: Stronger performance when analyzing multiple images simultaneously.
• Function Calling: Significant gains in its ability to call tools in both text and vision-text scenarios.

Benchmark results show LFM2.5-VL-3B performing well in its size class, particularly in grounding (RefCOCO-avg score of 87.9) and screen understanding (ScreenSpot-v2 scores above 78). These scores indicate a strong competency against other models like gemma-4-E4B-it and Qwen3.5-4B in these specialized, real-world tasks.

Implications for the Edge AI Market

By optimizing for on-device inference, LiquidAI is addressing the market for AI applications that require low latency and data privacy. The model's reported speed—achieving 228 tokens/s on an M5 Max and 20 tokens/s on a Galaxy S26 Ultra—makes it a viable option for developers building intelligent features into consumer electronics and desktop software. Its immediate availability on Hugging Face with support for inference libraries like llama.cpp and ONNX facilitates broad adoption, potentially reducing developer reliance on larger, more expensive cloud-based APIs for vision tasks.

With the release of LFM2.5-VL-3B, LiquidAI is making a calculated move to capture the edge AI market by prioritizing specialized, high-speed capabilities like UI grounding and on-device function calling over raw performance on general benchmarks—a strategy that directly targets developers building real-world, interactive applications.