ModelQwen2-VL

Qwen2-VL-2B

A multimodal vision-language model released by Alibaba's Qwen team. It supports joint understanding of images and text, and the 2B variant is a lightweight version with a relatively small parameter count.

Timeline (last 90 days)

Oct 14

Taking Qwen2-VL-2B as an example, the combined strategy integrating hard negative mining, false-negative mitigation and hard batch construction improves performance on MMEB by about 6.0% on average.

Unverified50%

All Facts (1)

Unverified

Taking Qwen2-VL-2B as an example, the combined strategy integrating hard negative mining, false-negative mitigation and hard batch construction improves performance on MMEB by about 6.0% on average.

50%

Source Articles

Contrastive Learning Strategies for Multimodal Embeddings: What Actually Works?Oct 14