LFM2.5-VL-3B-DragOn-GGUF

LFM2.5-VL-3B-DragOn — GGUF

GGUF quants of EryriLabs/LFM2.5-VL-3B-DragOn: LiquidAI's LFM2.5-VL-3B fine-tuned for drag-and-drop grounding on GUI screenshots. Screenshot + instruction in, {"start":[x,y],"end":[x,y]} out (0-1000 normalised coordinates).

The short version of the story: the base model scores 0.7% on the DragOn public eval, this fine-tune scores 70.8% acc@5 (78.5% acc@10), and it cost about $36 to train. Details, per-domain numbers and caveats are in the main repo.

Files

You need TWO files: a main model quant plus the vision projector (mmproj).

Note that coordinates are a precision task, so if you see degraded accuracy at Q4, step up a quant before blaming the model.

Usage

llama-server -m LFM2.5-VL-3B-DragOn-Q6_K.gguf \
  --mmproj mmproj-LFM2.5-VL-3B-DragOn-F16.gguf \
  -ngl 99 -c 4096 --temp 0

Then send a chat completion with the image and this exact prompt shape (it's what the model was trained on):

This is a screenshot of a user interface. You must perform a DRAG action.
Task: <your instruction>
Give the drag as JSON with the START point (where the mouse button goes down) and the END point (where it is released), in coordinates normalised to 0-1000 for both x (left->right) and y (top->bottom):
{"start":[x,y],"end":[x,y]}
Output only the JSON.

Multiply by your actual screen size /1000 and you have your drag.

Thanks

To LiquidAI for the base model and to Nathan Bout, Maxime Langevin and Ronan Riochet at Hcompany for the DragOn dataset — see the main repo card for the full credits.

Quantised by Dwain Barnes (EryriLabs), August 2026.