2B Models Handle Image Decisions as Same-Size General Models Fall Behind
In a set of 5060 Ti tests reposted by Liquid AI, two roughly 2B d1 models are compared with Qwen and InternVL on document classification, violence flagging, and part defect recognition. Latency is 0.13 seconds per item, with 4GB of VRAM. The tester also notes that comparing decision models with general chat models is not entirely apples-to-apples.
A single 5060 Ti, BF16 precision, two models with about 2B parameters, running three types of image judgments. Let's look at the data first.
The test comes from X user irony.aditya and was reposted by Liquid AI's official account. The models are Liquid's d1-2B and d1-Omni-2B, compared with Qwen3.5-2B and InternVL3.5-2B.
Task 1: route scanned documents into one of 16 types: d1-2B 88.8%, Qwen3.5-2B 55.0%, InternVL3.5-2B 56.2%.
Task 2: flag violent content: d1-Omni-2B 97.3%, Qwen 83.6%, InternVL 87.7%.
Task 3: identify defective parts: d1-2B 70.8%, Qwen 64.6%, InternVL 52.1%.
On speed, d1 takes 0.13 seconds per item, Qwen 0.26 seconds, and InternVL 0.64 seconds. The models have about 2B parameters and use 4GB of GPU VRAM. The tester wrote in the summary that d1 ranked first in every round.
The video demo page shows four models processing side by side, with 80 images waiting to be processed under each model. Task options include Route a document, Inspect a part, and Flag violence.
Fairness also needs to be made clear. The tester himself notes that d1 is a decision model, while the other two are general chat models, so strictly speaking they are not fully comparable. This caveat matters. It shows that what the data can demonstrate is: on fixed image-judgment tasks, specialized small models have an advantage; this cannot be directly extrapolated to mean stronger general capabilities.
What is more noteworthy about this set of results is local deployment. 4GB of VRAM, 2B parameters, and latency in the hundred-millisecond range mean that many fixed judgment tasks can be run locally without sending images to the cloud. But before choosing a model, you should still retest with your own data, especially class distribution, image quality, and the cost of misjudgment. A general model being strong at chat does not mean it is strong at judgment tasks; a specialized model being accurate at judgment does not mean it can chat. The two solve different problems.
Original tests and data can be found in Liquid AI's repost: https://x.com/liquidai/status/2116866695306625534
发布时间: 2026-11-02 03:44