Qwen2 VL 2B Instruct

Qwen2-VL-2B-Instruct is a vision-language model that processes images, videos, and text to perform visual understanding, question answering, document analysis, and reasoning tasks. The model introduces Naive Dynamic Resolution, allowing it to handle arbitrary image resolutions by mapping them into a variable number of visual tokens. It also employs Multimodal Rotary Position Embedding (M-RoPE), which separately captures positional information for 1D text, 2D images, and 3D video, improving multimodal processing. Key capabilities include understanding videos longer than 20 minutes for question answering and content creation, supporting text extraction in multiple languages (European languages, Japanese, Korean, Arabic, Vietnamese) within images, and operating as an agent for mobile and robotic devices via visual environment interpretation and text instructions.

Performance Benchmarks

The model achieves strong results across visual understanding benchmarks. Selected image and video benchmark scores are shown below.

Benchmark	Score
MMMU	41.1
DocVQA	90.1
InfoVQA	65.5
ChartQA	73.5
TextVQA	79.7
OCRBench	794
RealWorldQA	62.9
MMVet	49.5
MathVista	43.0
MVBench	63.2
Video-MME (wo/ subs)	55.6
Video-MME (w/ subs)	60.4

Qwen2-VL-2B-Instruct outperforms comparable models (e.g., InternVL2-2B, MiniCPM-V 2.0) on most reported benchmarks. This model is hosted by gigarouter as a managed, OpenAI-compatible API.