# Qwen3-VL-2B on standard vision tasks — retained record

> Retained reddit record from reddit:computervision.

Canonical: https://highsignal.app/data/records/30ed0c2161bed3a4

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:computervision)  published 2026-11-13 04:17 ingested 2026-11-14 02:36

[original source ↗](https://www.reddit.com/r/computervision/comments/1xuy7cj/qwen3vl2b_on_standard_vision_tasks/)

## retained text



```
Recently there has been lots of posts regarding the newer foundational models like LLM/VLMs and more traditional models.

So I thought it would be interesting to benchmark VLMs like **Qwen3-VL-2B** on more traditional vision tasks like classification and detection.

* **Classification - surprisingly good:** 93% top-1 accuracy, compared with 96% for Qwen3-VL-235B on the same samples. That makes the 2B model interesting for local classification.
* **Attributes - acceptable, but uses far more tokens:** 55.27% accuracy on OVAD, averaging 152 output tokens per sample versus 7 for classification—about 21× as many. Input token usage was also roughly 5× higher.
* **Detection - poor:** Only 6.75% recall. Additionally, 21% of samples failed to produce usable output, and 13% reached the token cap. I’d definitely keep an output limit in place.
* **Grounding - also weak:** 35% point-in-box accuracy and 22% box accuracy.

>Tl;DR Classification was the strongest result here; reliable localization was much harder.

Benchmark results, methodology and code:  
[https://github.com/Tenkei/vlm-tasks/tree/main/benchmarks](https://github.com/Tenkei/vlm-tasks/tree/main/benchmarks)
```



document fetched 2026-11-14 00:23 · canonical https://reddit.com/r/computervision/comments/1xuy7cj/qwen3vl2b_on_standard_vision_tasks
