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

Liquid AI has officially unveiled LFM2.5-VL-3B, its most sophisticated vision-language model (VLM) designed specifically for local deployment. By prioritizing direct, low-latency responses over complex reasoning chains, this 3.1-billion parameter model is engineered to thrive on edge hardware, from high-end workstations to mobile devices, without sacrificing accuracy or utility.

A Leap Forward in Edge Intelligence

The LFM2.5-VL-3B model represents a significant evolution in Liquid AI’s product lineup, introducing four critical enhancements that broaden its practical application:

• Advanced Screen and UI Comprehension: The model demonstrates a refined ability to interpret digital interfaces across a diverse range of device form factors.
• Enhanced Grounding: Object detection and spatial awareness have been bolstered, allowing the model to respond accurately to natural language queries regarding specific visual elements.
• Multi-Image Reasoning: The architecture now supports more robust analysis when processing multiple images simultaneously.
• Superior Function Calling: Whether operating in text-only or multimodal environments, the model shows a marked improvement in executing external tools and functions.

"LFM2.5-VL-3B is our most capable vision-language model you can run on your own hardware. It understands documents and screens alike, grounds objects, and can call tools. It answers directly instead of reasoning, so responses stay fast in real-time and on-device apps."

The Architecture Behind the Performance

To achieve this level of efficiency, Liquid AI paired a SigLIP2 400M NaFlex vision encoder with the robust backbone of the LFM2.5-2.6B text model. The training regimen was extensive, utilizing approximately 34 trillion tokens. Notably, the team increased the volume of vision-centric data by four times compared to previous iterations, incorporating a curated mix of synthetic image-caption pairs, OCR datasets, grounding tasks, and instruction-following sets.

Recognizing the global nature of modern AI applications, the developers expanded the model's vocabulary to 128,000 tokens. By extending the tokenizer in place rather than starting from scratch, they ensured the model maintains high performance across non-Latin scripts. The post-training process followed a two-tiered approach: an initial stage of supervised fine-tuning (SFT) bolstered by knowledge distillation from a larger teacher model, followed by a multi-reward reinforcement learning (RL) phase to refine output quality.

Benchmark Dominance

In rigorous testing, LFM2.5-VL-3B consistently outperformed its peers in its size class. The evaluation covered a broad spectrum of tasks, including multilingual visual comprehension, scientific reasoning, document analysis, and complex GUI navigation.

Key Performance Metrics

The model’s performance is particularly notable in real-world scenarios. On the MMStar benchmark, it achieved a score of 63.3, and it demonstrated exceptional proficiency in document-based tasks, scoring 91.1 on DocVQA. Furthermore, its ability to handle GUI-related tasks—as evidenced by its high scores on the ScreenSpot-v2 benchmarks—positions it as a premier choice for automation and assistive technology.

When evaluated on text-only instruction following and tool-use benchmarks, the model proved it is not just a vision specialist. It stands on par with larger models like the Gemma-4-E2B and Qwen3.5-2B, proving that a smaller footprint does not necessitate a compromise in logical capability or functional versatility.

Unmatched Inference Speed

Efficiency is the hallmark of LFM2.5-VL-3B. The model is optimized for immediate integration into the existing inference ecosystem, including llama.cpp, MLX, vLLM, SGLang, and ONNX.

• On-Device Performance: On an M5 Max chip, the model achieves a decoding speed of 228 tokens per second, while maintaining 116 tokens per second on the Ryzen AI Max+ 395. Even on mobile hardware like the Galaxy S26 Ultra, it remains highly responsive, hitting 20 tokens per second.
• GPU Throughput: In data center environments, the model excels, reaching approximately 11,000 tokens per second at high concurrency. This represents a throughput nearly double that of 4B-class models, allowing for the generation of nearly 1 billion tokens per day on a single H100 GPU.

Getting Started with LFM2.5-VL-3B

Liquid AI has prioritized accessibility, ensuring developers can integrate this technology with minimal friction. The model is available now on Hugging Face. To begin, users can leverage the transformers library (version 5.10.1 or higher) to load the model:

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

MODEL_ID = "LiquidAI/LFM2.5-VL-3B"
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForImageTextToText.from_pretrained(
    MODEL_ID, device_map="auto", dtype="bfloat16"
)

For those who prefer a hands-on experience without local setup, Liquid AI provides a WebGPU-powered browser demo. This interface allows users to upload multiple images and interact with the model directly, showcasing its capabilities in OCR, grounding, and tool utilization in real-time.

Final Thoughts

With the release of LFM2.5-VL-3B, Liquid AI is delivering on its core mission: making high-performance AI ubiquitous. By balancing the demands of visual understanding with the constraints of edge hardware, the company has created a tool that is as practical as it is powerful. Whether you are building document-processing pipelines, mobile assistive apps, or automated UI agents, LFM2.5-VL-3B provides the speed and intelligence required to push the boundaries of what is possible on local hardware.

*

Citation: Liquid AI, "LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge", Liquid AI Blog, Aug 2026.