Qwen-VL-Chat
Alibaba · 2023-08-22
Qwen-VL-Chat is Alibaba Cloud's instruction-tuned multimodal assistant based on Qwen-VL, released August 22, 2023. It combines a Qwen-7B-initialized language model with a ViT-bigG-initialized visual encoder and a single-layer cross-attention adapter; the paper reports rounded component sizes of 7.7B, 1.9B, and 0.08B (9.6B total). It supports multi-image, multi-round multilingual dialogue, visual question answering, text reading, and text-encoded grounding; although the later multimodal training stages used 2,048-token LLM sequences, the shipped tokenizer and configuration declare an 8,192-token maximum with dynamic NTK and LogN scaling.
Benchmark scores
| Benchmark | Score |
|---|---|
| AI2D (OpenVLM) | 63.0 |
| Capability | 94.2 |
| HallusionBench (OpenVLM) | 36.8 |
| MathVista (OpenVLM) | 35.3 |
| MM-Vet (OpenVLM) | 47.3 |
| MMMU Validation (OpenVLM) | 37.0 |
| MMStar (OpenVLM) | 34.5 |
| Video-MME | 41.1 |