Qwen2.5 VL 3B

Qwen2.5-VL-3B-Instruct is a vision-language model (VLM) that processes images, videos, and text to perform visual understanding, document analysis, visual agent tasks, and structured output generation. It is the 3-billion-parameter instruction-tuned variant in the Qwen2.5-VL series.

Key Capabilities

• Visual understanding: Recognizes common objects, texts, charts, icons, graphics, and layouts within images.
• Agentic functionality: Acts as a visual agent capable of reasoning and dynamically directing tools, including computer and phone use.
• Long video comprehension: Understands videos over one hour and can pinpoint relevant video segments for specific events.
• Visual localization: Generates bounding boxes or points for objects and provides stable JSON outputs for coordinates and attributes.
• Structured outputs: Extracts structured data from invoices, forms, tables, and similar documents.

Architecture

Employs a dynamic resolution mechanism for images and dynamic FPS sampling for videos, with an updated vision encoder using window attention, SwiGLU, and RMSNorm. Multimodal Rotary Position Embedding (M-RoPE) enables fusion of positional information across text, images, and videos.

Benchmark Performance

Image benchmarks:

Benchmark	Qwen2.5-VL-3B	Qwen2-VL-7B	InternVL2.5-4B
MMMU	53.1	54.1	52.3
DocVQA	93.9	94.5	91.6
InfoVQA	77.1	76.5	72.1
MathVista	62.3	58.2	60.5
MathVision	21.2	16.3	20.9

Video benchmarks: MLVU: 68.2, VideoMME: 67.6/61.5, MVBench: 67.0, EgoSchema: 64.8, PerceptionTest: 66.9.

Agent benchmarks: ScreenSpot: 55.5, AndroidWorld_SR: 90.8, AITZ_EM: 76.9.

Input/Output

Accepts interleaved images, videos, and text. Outputs text, bounding boxes, points, and structured JSON. Supports dynamic token allocation per image (4–16,384 visual tokens).