Liquid AI has released LFM2.5-VL-3B, a lightweight vision-language model, and it is worth your attention if you use AI Bridge to generate descriptions and keywords for the images in your catalog.
It is a 3-billion-parameter vision model that outperforms models more than twice its size, and it runs remarkably fast. We tested it in the usual AI Bridge + LM Studio setup, and at 4-bit quantization (Q4_K_M) it beat Gemma 4, the model we currently recommend as the default, on both description quality and speed.

What we saw in testing
Q4_K_M files take about 2.5 GB, so it fits comfortably on machines where larger vision models are a stretch.As always with local models, everything stays on your computer: no images are uploaded anywhere, and there are no subscriptions or per-image fees.
How to use it with AI Bridge
The setup is exactly the same as described in Getting Started with AI Bridge, with two differences:
LFM2.5-VL-3B instead of gemma-4 and pick LiquidAI/LFM2.5-VL-3B-GGUF.Q4_K_M quantization. It is enough — the heavier variants are not worth the extra disk space and memory for this task.Then load the model, start the local server, and select the model in AI Bridge Options as usual.
A tip: get brand names into your keywords
There is a Custom Instructions field in AI Bridge Options, and whatever you type there is added to every request. LFM2.5-VL-3B is good at recognizing logos and product packaging, so it is worth telling it to put that information where you can search for it:
If you recognize well-known brands, logos, or product models in the image, include their names in the keywords.
With that in place, recognized brand names end up in the keyword list for the image, and become searchable in WinCatalog along with everything else AI Bridge generates. Handy for product shots, packaging photos, car pictures, and screenshots of software.
Bottom line
Go ahead and download LFM2.5-VL-3B from Hugging Face (or straight from the LM Studio model search) and use it in AI Bridge. On our test images it is the best combination of quality, speed, and size we have tried so far.
Posted on: August 20, 2026