lfm2.5-vl-3b vs qwen3.5-4b: 3b beats 4.2b on visual grounding, but neither finds the button

• lfm2.5-vl-3b – liquid ai, 3b params, lfm2.5-2.6b backbone + siglip2 400m vision encoder

• qwen3.5-4b – alibaba, 4.2b params, unsloth gguf build

we put the two models on one job: point to the right spot in a screenshot, a receipt, and a photo

the setup: both run locally, same q5_k_m quantization, one bounding-box prompt per task, scored by hand against a pixel-checked ground truth box

- task 1. name the object closest to the camera in a desk-clutter photo
- task 2. find "add to cart" on an amazon product page, two buttons stacked 9px apart
- task 3. read the total off a photographed receipt, total sits between subtotal and debit tend

1. octopus
• lfm2.5-vl-3b – hit, 8.5s, 21 tokens
• qwen3.5-4b – hit, 6.9s, 352 tokens

2. button
• qwen3.5-4b – close miss, boxed "buy now" instead of "add to cart". 19.2s, 1.2k tokens
• lfm2.5-vl-3b – missed entirely, boxed half the product description. 1.8s, 13 tokens

3. receipt
• lfm2.5-vl-3b – landed on "46.42", the receipt's total. counted as a hit
• qwen3.5-4b – missed, boxed the "debit" line

observations:

- qwen's button miss was a bullseye on the wrong target

– tight box on "buy now", 9px from the button we asked for

- on the button task qwen took 19.2s and 1.2k tokens to land on the wrong element – lfm answered in 1.8s on 13 tokens, also wrong, but 10x faster to be wrong

- lfm never found the button in three tries – share icon, a stale duplicate answer, then half the product description boxed as one region

- on the receipt both models skip "total" as a label – lfm lands on the right number one row below, qwen jumps to a different line entirely

- octopus is the only clean tie, near-identical boxes from both models

- qwen spent 352 tokens grounding the octopus, 16x lfm's 21, for the same answer

- we also tried gemma-4-e4b – it described the image fine but refused to return a bounding box on every attempt

conclusion: 3b beat 4.2b on tokens per call and on the receipt, tied on the photo, and neither model boxed the right button once!

watch the full test via link