On-Device Vision AI Reads Screens in 3GB of Memory

A Vision Model That Fits Where the Data Already Is

Most vision-language models still assume a round trip to a data center. You photograph a receipt, a form, or a screen, and the pixels travel to someone else's GPU before an answer comes back. On August 12, 2026, Liquid AI released LFM2.5-VL-3B, a 3.1-billion-parameter vision-language model built on the opposite assumption: that the interesting images are already on the device, and so the model should be too.

• 3.1 billion parameters running in roughly 3GB of memory, small enough for phones, laptops, and single-board machines
• 69.4% average across vision and text benchmarks, including 81% on MMBench and 91.1% on document visual question answering
• 228 tokens per second decoding on an Apple M5 Max, 116 tokens/second on Ryzen AI, and about 20 tokens/second on a Galaxy S26 Ultra
• Trained on 34 trillion tokens with roughly four times more vision data than its predecessor

The headline is not that a small model scored well. It is that the capability list — screen understanding, object grounding, multi-image input, and function calling — is the list you would want from a cloud model, and it runs locally.

What Can a 3B Vision Model Actually Do?

The four capabilities Liquid AI highlights are worth separating, because they compose into something more useful than image captioning.

Screen and interface understanding means the model can look at a rendered application, not just a photograph, and make sense of the layout. Grounding means it can point — given a description, it returns where in the image that thing is, scoring 87.9% on object grounding tasks. Multi-image input lets it reason across a sequence rather than a single frame. Function calling means the output can be a structured tool call instead of prose.

Put together, that is the raw material for an agent that can see. A model that can read a screen, locate a control on it, and emit a tool call is doing the perception half of computer use. Liquid AI is explicit that the model is tuned to answer quickly rather than to produce long reasoning chains, which is the right trade for interface work where latency is the whole experience.

Document handling is the other obvious fit. A 91.1% score on document visual question answering, running on the machine that holds the documents, is a meaningfully different privacy posture than uploading them. This is the same on-device direction we have tracked across our edge AI and on-device model coverage, and it keeps arriving faster than expected.

How Does It Run So Fast on Small Hardware?

The architecture pairs a SigLIP2 vision encoder with a pre-trained language backbone, and the efficiency story shows up most clearly on throughput rather than parameter count. On GPU infrastructure the model reaches roughly 11,000 output tokens per second, which Liquid AI reports as about double what comparable 4B-class models manage.

Deployment support is broad enough to matter: native paths through llama.cpp, MLX, vLLM, and ONNX cover most of how people actually run local models today. That breadth is the difference between a model you can benchmark and a model you can ship, and it follows the same practical instinct as the company's earlier LFM2.5-2.6B tool-calling release and its CPU-optimized encoder work.

Why Small Multimodal Models Matter Now

There is a pattern forming across the open-weight field, and LFM2.5-VL-3B sits squarely in it. Capability that required a frontier model eighteen months ago is being compressed into something that fits a laptop, and the compression is not costing as much accuracy as anyone expected. NVIDIA's Nemotron 3.5 Lightning made the same argument for long-running text agents; this makes it for perception.

For developers, the practical consequence is that the default architecture for a vision feature is changing. The question is no longer whether a local model is good enough, but whether the workload genuinely needs a cloud round trip at all. For a receipt scanner, an accessibility tool, or an interface agent, the answer is increasingly no — and the version that runs locally is the one that works on a plane, in a hospital basement, and on a device that never uploads the picture.

Sources: Liquid AI on Hugging Face — August 12, 2026; MarkTechPost — August 13, 2026.