Liquid AI launches LFM2.5-VL-3B for edge vision

Liquid AI has launched LFM2.5-VL-3B, an AI model that can understand images, documents, and screens, designed to run directly on devices with limited resources. The company says it can operate with around 3 GB of memory and reach up to 20 tokens per second on a Galaxy S26 Ultra.

The model belongs to the vision-language model category: systems that combine text and images to answer questions, read documents, or act on an interface. It has 3.1 billion parameters, far fewer than many cloud-hosted multimodal models.

What it can do

Liquid AI highlights four improvements over its previous model, LFM2-VL-3B:

• Understand screens and applications on computers, phones, and web pages.
• Locate objects based on natural-language instructions, such as pointing out where a cup or button appears.
• Reason over multiple images at the same time, for example to compare products or track changes between photographs.
• Call tools through structured instructions, both when it receives text and when it analyzes images.

This makes it useful for specific tasks: extracting data from an invoice, identifying an element in a screenshot, answering questions about a chart, or opening an application feature after interpreting what appears on screen.

A model designed to run without the cloud

LFM2.5-VL-3B combines a 400-million-parameter SigLIP2 visual encoder with the LFM2.5-2.6B text backbone. Liquid AI says it trained the model on around 34 trillion tokens and with four times more visual data than in its previous generation.

The training set includes images with captions, text extracted from documents, object-localization examples, and synthetic instructions. The company also expanded its vocabulary to 128,000 units to improve reading in languages that use non-Latin alphabets, without retraining the tokenizer from scratch.

In its internal tests, the model achieved an average score of 69.4 out of 100 across the group of visual evaluations it compares with other small models. It does not lead every test, but stands out especially in object localization and interface understanding.

For example, it scored 87.9 on RefCOCO-avg, a text-guided object-localization test. On ScreenSpot-v2, which measures whether a model can identify interface elements, it scored 78.7 on desktop, 81.2 on mobile, and 82.2 on web.

These figures come from the evaluations published by Liquid AI. They were conducted without enabling reasoning modes and with the generation parameters recommended for each model, so they do not necessarily represent performance across all real-world applications.

Tool use also improves

The ability to call functions matters because it allows a model to do more than describe what it sees. It can return a structured instruction to search for a product, query a database, or perform an action in an application.

In the ToolSandbox test, LFM2.5-VL-3B scored 59.5, compared with 26.4 for its previous version. On BFCL V4, it reached 32.5. According to Liquid AI, its tool-use results are close to models such as Gemma-4-E2B and Qwen3.5-2B, although results vary by test.

Speed and availability

On the hardware tested by the company, the model generates up to 228 tokens per second on an M5 Max and 116 on a Ryzen AI Max+ 395. On servers handling many simultaneous requests, it reached around 11,000 tokens per second, approximately twice the speed of some models in the 4-billion-parameter class.

The model supports llama.cpp, MLX, vLLM, SGLang, and ONNX from launch. You can also download it from Hugging Face, try it in a WebGPU demo in your browser, and adapt it through fine-tuning.

For you, the main difference is that an AI capable of reading images and screens can run locally, with less dependence on an internet connection and greater control over your data. The next thing to watch is whether that performance holds outside the manufacturer's tests, especially on phones and in tasks that require reliable action inside real applications.