Liquid AI has released LFM2.5-VL-DSpark, an experimental add-on for its LFM2.5-VL-3B model, which works with both images and text. The announcement is about making answers arrive faster. It does not give the main model a new ability to understand images.
What does that mean?
Imagine an AI being asked what is in a photograph or to read a scanned document. After it processes the image, it has to produce an answer piece by piece. DSpark acts as a small helper: it proposes several pieces of the answer, and the main model checks them. When enough proposals are accepted, the answer can be produced more quickly without changing which model makes the final decision.
What is it used for?
The underlying LFM2.5-VL-3B model is available to developers for tasks such as describing images, reading documents and screens, identifying objects, and answering questions about charts. Liquid AI offers downloads and a browser demonstration of the main model. DSpark is a new, experimental way for developers using supported software to speed up its answers; Liquid AI has not identified a consumer app or customer deployment already using this add-on.
In Liquid AI’s own tests on six visual tasks, the answer-generation stage on an M5 Max was 2.30 to 3.13 times faster. The improvement to the whole response was smaller: 1.56 to 2.62 times faster. Those are company-reported results on specified hardware and settings, not a promise that every app or device will see the same gain. DSpark does not speed up the earlier work of processing the image or the prompt.
For readers, the practical possibility is a more responsive visual assistant in software that adopts the add-on—for example, one that describes a photo or extracts information from a document. Nothing in the announcement shows that existing apps have switched to it, and there is no reason for an ordinary user to download DSpark on its own.
Sources: Liquid AI’s DSpark announcement and the underlying model card.