What if a model could read documents, understand a screen, and analyze multiple images without sending your data to the cloud? Liquid AI presents LFM2.5-VL-3B, a vision and language model designed to perform multimodal tasks quickly on local devices and low-power hardware.
With 3.1 billion parameters, the model aims for a rare balance: advanced visual capabilities, tool support, and a small enough memory footprint to run even on compatible phones.
Four improvements for understanding images and screens
LFM2.5-VL-3B expands the capabilities of its predecessor, LFM2-VL-3B, in four main areas:
- Screen and interface understanding: interprets elements in mobile apps, websites, and desktop environments.
- Grounding: locates objects in an image based on natural-language instructions.
- Multiple-image input: compares and reasons about several images within the same query.
- Function calling: can decide when to use a tool, both in text conversations and in scenarios that combine images and text.
In practice, you could show it a screenshot and ask it to identify the button for downloading a file. You could also give it several photos of a product, have it analyze a receipt, and ask it to extract the data so you can register it in another application.
The important difference is not just describing an image. It is connecting what appears in it to a concrete action.
A compact, but multimodal architecture
The model combines a 400-million-parameter SigLIP2 vision encoder with NaFlex and the same pretrained backbone as the LFM2.5-2.6B text model. This combination makes it possible to process visual and textual information within a relatively small system.
Liquid AI says training used approximately 34 trillion tokens and four times more visual data than the previous generation. The dataset included image-description pairs, OCR data, grounding examples, and synthetic and curated instructions.
The team also expanded the tokenizer's vocabulary to 128,000 tokens to improve support for non-Latin scripts. Instead of training from scratch, it extended the existing tokenizer—a decision that reduces the cost of adapting the model to more languages and writing systems.
Training in two stages
Post-training took place in two phases. First, the team applied supervised fine-tuning, known as SFT, with knowledge distillation from a larger teacher model and a training method called Antidoom.
Then it used reinforcement learning with multiple rewards. Put simply, the model did not just learn correct examples; it was also optimized to follow instructions, detect objects, use tools, and produce more useful responses.
Notable results in vision and language
Liquid AI evaluated LFM2.5-VL-3B on tests covering visual understanding, OCR, documents, mathematics, scientific reasoning, object detection, multiple images, and graphical user interface understanding.
In the comparison presented by the company, the model achieved an average of 69.4 points across the visual test suite. It matched the average score of InternVL 3.5 4B and ranked above several similarly sized models, although Qwen3.5-4B reached 70.1 points.
Some results stand out in particular:
- RefCOCO-avg: 87.9 points in grounding, compared with 57.1 for LFM2-VL-3B.
- BLINK: 61.5 points in multiple-image reasoning, compared with 50.2 for the previous generation.
- MuirBench: 58.3 points, a jump from the previous 34.9.
- TextVQA: 84.3 points on questions about text within images.
- DocVQA: 91.1 points in document understanding.
- ScreenSpot-v2: between 78.7 and 82.2 points, depending on whether the task involved desktop, mobile, or web interfaces.
These figures need to be interpreted in context. The tests were run with vLLM 0.26.0, using each model's recommended parameters whenever they were available. The evaluation also used the mode without explicit reasoning and instructed the models to answer directly.
In other words, the results do not represent every possible situation. They do show, however, that a compact model can be competitive on tasks that typically require much larger systems.
Better tool and function use
The improvement is not limited to images. In instruction-following benchmarks, LFM2.5-VL-3B reached 82.3 points on IFEval, compared with 72.9 for LFM2-VL-3B.
For tool use, it scored 59.5 points on ToolSandbox, compared with 26.4 for the previous generation. On BFCL V4, it achieved 32.5 points, above LFM2-VL-3B's 20.5.
Liquid AI says the model is on par with Gemma-4-E2B and Qwen3.5-2B for tool use. This opens the door to assistants that do more than answer questions: they can also extract information from an image, select a function, and return structured data to another application.
Speed for running on the device
Liquid AI's approach is geared toward what's known as edge AI: running artificial intelligence close to the user, without necessarily depending on a remote server.
According to the published measurements, LFM2.5-VL-3B reaches:
- 228 tokens per second on a computer with an M5 Max.
- 116 tokens per second on a Ryzen AI Max+ 395.
- Around 20 tokens per second on a Galaxy S26 Ultra.
- Approximately 3 GB of memory for local execution.
On GPUs, the company also presents it as the fastest model in its evaluation for multi-frame inputs. With high concurrency, it reaches about 11,000 output tokens per second—approximately twice as fast as models in the 4-billion-parameter class.
For a company, this can translate into lower latency and reduced infrastructure costs. For a user, it could mean that an application analyzes a screenshot, document, or photograph without uploading it to an external service.
Available across several inference environments
LFM2.5-VL-3B supports several execution tools from day one:
llama.cppMLXvLLMSGLangONNX
It can also be loaded with Transformers. A basic workflow uses AutoProcessor to prepare the image and AutoModelForImageTextToText to generate the response. The example published by Liquid AI shows how to ask the model to describe an image in two sentences.
The demo output correctly identifies two sleeping cats on a pink sofa next to two remote controls. It may sound like a simple task, but it helps verify the complete workflow: loading an image, processing it, and generating a multimodal response.
What it means for developers and users
For developers, the appeal lies in building visual applications with more modest hardware requirements. A document assistant, an accessibility tool, or a technical support system could process images locally and trigger functions through structured calls.
For users, the clearest advantage is the potential for greater privacy. An application that understands a screen or document without sending it to the cloud can reduce the exposure of personal information. That said, actual privacy will depend on how each application implements the model, what data it stores, and what permissions it requests.
LFM2.5-VL-3B does not eliminate the need for large models. For complex tasks, extended reasoning, or very large contexts, bigger systems may still have an advantage. Its proposal is different: provide sufficient visual and tool-use capabilities with low latency and local execution.
The trend is clear. Visual AI is no longer an exclusive feature of data centers and is beginning to reach personal computers, phones, and specialized devices. The question is no longer whether a machine can see, but what useful tasks it can complete with that information without forcing you to depend on the cloud all the time.