LFM2.5-VL-3B: The Tiny Multimodal Model Built for AI That Runs Anywhere

The most interesting AI models are not always the ones with hundreds of billions of parameters. Sometimes the real breakthrough is making a surprisingly capable model small enough to run almost anywhere.

Liquid AI’s LFM2.5-VL-3B is built around exactly that idea. It is a multimodal vision-language model designed for on-device deployment, combining a 2.6 billion parameter LFM2.5 language backbone with a 400 million parameter SigLIP2 NaFlex vision encoder. It can understand both text and images while targeting low-latency workloads on everything from desktop hardware to smartphones.

Small Enough to Actually Run Locally

LFM2.5-VL-3B is reported at roughly 3.1 billion total parameters, but its practical hardware requirements are what make it interesting.

Liquid AI reports that the model can decode around 228 tokens per second on an Apple M5 Max and 116 tokens per second on an AMD Ryzen AI Max+ 395, while fitting in roughly 3 GB of memory. The company also reports around 20 tokens per second on a Galaxy S26 Ultra.

That changes the conversation around local multimodal AI.

A model that can look at photographs, screenshots, documents, charts, signs, and interfaces does not necessarily need to live behind a cloud API. With LFM2.5-VL-3B, developers can build applications where the image stays on the user’s device and inference happens locally.

For privacy-sensitive applications, that matters just as much as raw benchmark performance.

More Than Image Captioning

At first glance, a 3B vision-language model might sound like another image-description system. LFM2.5-VL-3B goes considerably further.

Liquid AI specifically highlights improved visual grounding and object detection through natural-language queries, as well as full-page OCR with layout annotations. Its layout system can identify regions such as titles, paragraphs, tables, charts, equations, code, headers, footnotes, and images while returning their coordinates and extracted content.

That makes it particularly useful for document-heavy applications.

Imagine pointing your phone at an invoice and extracting the relevant fields. Or scanning a restaurant menu and translating it. Or processing thousands of scanned pages and converting them into structured, searchable text. These are exactly the kinds of workloads Liquid AI recommends for the model.

The 32K Context Is Deliberate

LFM2.5-VL-3B has a 32,768-token context window. That is nowhere near the enormous context windows advertised by some larger cloud models, but that is not really the point of this model.

Liquid AI explicitly recommends LFM2.5-VL-3B for single-turn, high-throughput and low-latency workloads rather than long-context, reasoning-heavy tasks.

That distinction is important.

This is not positioned as a replacement for a giant frontier reasoning model that you feed an entire software repository. It is designed more like a specialized visual intelligence engine: give it an image or document, ask a focused question, and get a fast answer.

That specialization is one of its strengths.

The Architecture Behind the Speed

The model combines Liquid AI’s LFM2.5-2.6B language model with a shape-optimized 400M SigLIP2 NaFlex vision encoder. NaFlex allows the model to work with variable image layouts and native resolutions rather than forcing every image into a single rigid shape. Large images can be processed as non-overlapping 512×512 patches alongside a resized whole-image thumbnail.

The result is a multimodal architecture designed around efficiency rather than brute-force scale.

There are also several deployment formats available. Developers can use the native Hugging Face checkpoint with Transformers, vLLM, or SGLang, while GGUF exports target llama.cpp and ONNX exports support cross-platform deployment. Apple Silicon users can use the MLX version for optimized local inference.

That ecosystem makes the model considerably easier to integrate into existing local-AI workflows.

The Benchmark Numbers Are Surprisingly Strong

The model’s performance is particularly interesting because of its size.

On Liquid AI’s reported evaluations, LFM2.5-VL-3B scores 80.7 on ScreenSpot-v2, 87.9 on RefCOCO, 61.5 on BLINK, 58.3 on MuirBench, and 59.5 on ToolSandbox. It also reaches 81.3 on ChartQA, 68.5 on MathVista, and 47.5 on OCRBench v2.

These numbers do not make it universally better than larger models. For example, some 4B-class competitors outperform it on individual evaluations such as MMStar, OCRBench, and MMMU Pro. But the broader story is impressive: Liquid AI is achieving useful multimodal capabilities inside a model intended to run with only a few gigabytes of memory.

The key metric here is not simply “does it win every benchmark?”

It is “how much capability can you get per unit of hardware?”

And that is where LFM2.5-VL-3B becomes compelling.

A Model for Real-World Edge Applications

The strongest use cases are not necessarily chatbots.

LFM2.5-VL-3B makes sense for applications that continuously need to understand visual information without sending every frame or document to a remote server.

Think of smart cameras, document scanners, point-of-sale systems, industrial inspection, accessibility tools, vehicle interfaces, translation applications, and local productivity software.

A camera application could ask the model to identify a particular object. A document pipeline could OCR a page while preserving its layout. A mobile application could translate a road sign without an internet connection. Liquid AI also demonstrates browser and webcam-based experiences where users can interact with images and receive visual grounding information such as bounding boxes.

This is the direction edge AI has been moving toward: models that are not merely small versions of cloud systems, but deliberately engineered around the constraints of real devices.

But There Is an Important Catch

There is a temptation to read “3B vision-language model” and assume that it can replace every larger multimodal model.

It cannot.

Liquid AI explicitly recommends against using LFM2.5-VL-3B for long-context, reasoning-intensive tasks such as visual web design or highly technical questions about complex blueprints.

That limitation is important.

The model is optimized for fast, focused inference rather than deep, extended reasoning. It is much closer to a highly efficient visual utility model than an all-purpose frontier intelligence system.

And that is actually a useful distinction when choosing an AI model.

Why This Model Matters

The most important part of LFM2.5-VL-3B may not be any individual benchmark.

It is the hardware requirement.

Liquid AI is demonstrating that multimodal AI does not always require a massive GPU server. A model capable of understanding images, extracting document structure, detecting objects from language queries, and supporting tool-oriented workflows can fit into a memory footprint of only a few gigabytes and run directly on consumer hardware.

That opens the door to applications where latency, privacy, offline availability, or operating cost matter more than having the absolute largest model available.

Cloud AI is still powerful.

But edge AI is becoming practical enough that developers can start designing applications around a different assumption:

The AI does not have to live in a data center.

It can live inside the device.

What Developers Can Do With LFM2.5-VL-3B

1. Build local document scanners: Use its OCR and layout understanding to turn invoices, forms, and PDFs into structured data.
2. Create offline visual assistants: Let a phone or laptop answer questions about photographs, objects, signs, and screenshots without sending the images to a server.
3. Add visual grounding to applications: Natural-language queries can be combined with object detection and bounding-box style outputs for more interactive interfaces.
4. Run multimodal AI on modest hardware: The reported memory footprint makes experimentation possible on hardware far below the requirements of large vision-language models.
5. Deploy through existing AI infrastructure: Transformers, vLLM, SGLang, llama.cpp, ONNX, and MLX are all supported through different model variants and integrations.

The important lesson from LFM2.5-VL-3B is not that small models have suddenly defeated large models.

It is that the boundary between “AI that runs in the cloud” and “AI that runs on the device” keeps moving.

A 3B multimodal model that can process documents, understand screens, recognize objects, and generate useful responses at high speed is already enough to build genuinely useful products.

For developers interested in local AI, that may be more important than another record-setting benchmark.