Liquid AI releases LFM2.5-VL-3B, an open-weight vision model built to run on-device
Originally announced by Liquid AI . We link the primary source so you can read it for yourself.
Liquid AI released LFM2.5-VL-3B on 12 August 2026, a 3.1-billion-parameter open-weight vision-language model aimed at running on-device rather than in the cloud. The design goal is edge deployment: phones, laptops, and single consumer GPUs. The weights are on Hugging Face with day-one exports in GGUF, ONNX, and MLX formats, and Liquid AI lists support across llama.cpp, MLX, vLLM, SGLang, and ONNX runtimes from launch. It extends the company’s earlier LFM2-VL-3B with, Liquid AI says, stronger screen understanding, object grounding, multi-image reasoning, and function calling.
The licence, read carefully
The model is open-weight but not unconditionally open for commercial use. It is released under the LFM Open License v1.0, which Liquid AI describes as permitting commercial use free of charge only for organisations under 10 million US dollars in annual revenue; above that threshold a separate commercial arrangement is required. That is a meaningful condition rather than a formality, so if you are deploying this in a business, read the licence text and check where your organisation sits before you rely on it. It is more permissive than a research-only release and more restrictive than a straightforward Apache or MIT licence, which is a distinction the catalogue records precisely because it changes who can use a model and how.
Why an on-device VLM matters
Most of the local AI conversation is about text models. A capable vision-language model small enough to run entirely on a phone or a laptop is a different and genuinely useful thing: it means image and screen understanding, document reading, and visual question answering can happen without sending a photograph or a screenshot to anyone else’s server. For privacy-sensitive uses, and for anything that has to work offline, that is the whole point of running locally, extended to images. The function-calling and screen-understanding features point at the obvious application, which is on-device agents that can see an interface and use tools against it.
At 3.1 billion parameters, this sits at a size that runs comfortably on modest hardware. As a rough guide a model this size needs in the region of 2GB to 4GB of memory at a 4-bit quantisation, which is well within a mid-range phone or any recent laptop, though the exact figure depends on the build and the image resolution you feed it. Treat that as an estimate rather than an official requirement.
The performance claims
Liquid AI positions the model as competitive with vision-language models around twice its size while running faster, and cites on-device speed of roughly 228 tokens per second on an Apple M5 Max. As with any figure from the releasing organisation, treat these as claims pending independent testing: a throughput number in particular depends heavily on the hardware, the runtime, the quantisation, and the size of the image being processed, so your own numbers may differ substantially. The M5 Max reference is a neat coincidence of timing, landing in the same window as Apple’s M5 Mac Studio and Mac mini refresh, though this model is squarely aimed at far smaller and cheaper hardware than that.
Where it sits
Liquid AI’s LFM line uses the company’s own architecture rather than a standard transformer, which is part of how it targets efficiency at small sizes. In the open, locally-runnable vision-language space the familiar comparison points are models like Qwen2.5-VL 7B, which is larger, and the smaller VLMs that have been appearing specifically for edge use. The trend this release fits is a clear one: vision-language models are getting small and fast enough to run on the device in your hand, not merely on a workstation, and the tooling to run them locally is arriving on day one rather than months later.
Worth a look if
You want on-device image or screen understanding, value privacy or offline operation, and can work within the licence’s revenue condition. The broad day-one runtime support, especially the MLX and GGUF exports, makes it unusually easy to try on hardware you already own.