跳轉到主要內容

Qwen2.5 VL 72B Instruct

Alibaba Cloud / Qwen TeamQwenOpen Weighttongyi-qianwen

描述

Qwen2.5-VL is the new flagship vision-language model of Qwen, significantly improved from Qwen2-VL. It excels at recognizing objects, analyzing text/charts/layouts in images, acting as a visual agent, understanding long videos (over 1 hour) with event pinpointing, performing visual localization (bounding boxes/points), and generating structured outputs from documents.

發布日期
2025-01-26
參數規模
72.0B
上下文長度
32K
支援模態
image, text

能力雷達圖

50
general
0
coding
60
reasoning
60
science估算
40
agents
80
multimodal

Science 在缺少專門科學評測時使用推理能力代理估算。

排行榜排名

領域#排名分數來源
智能体与工具82
45.0
LS
多模态榜54
73.0
LS
推理能力74
55.0
LS

基準測試分數 (LLM Stats)

Agents

AITZ_EM83.2%自報
MobileMiniWob++_SR68.0%自報
AndroidWorld_SR35.0%自報
OSWorld8.8%自報

General

MMVet76.2%自報
MLVU-M74.6%自報
MMStar70.8%自報
MMMU70.2%自報
MMMU-Pro51.1%自報

Grounding

ScreenSpot87.1%自報
ScreenSpot Pro43.6%自報

Image To Text

DocVQA96.4%自報
OCRBench88.5%自報
OCRBench-V2 (en)61.5%自報

Long Context

EgoSchema76.2%自報
LVBench47.3%自報

Math

MathVista-Mini74.8%自報
MathVision38.1%自報

Multimodal

Android Control Low_EM93.7%自報
ChartQA89.5%自報
AI2D88.4%自報
MMBench88.0%自報
CC-OCR79.8%自報
TempCompass74.8%自報
VideoMME w/o sub.73.3%自報
PerceptionTest73.2%自報
MVBench70.4%自報
Android Control High_EM67.4%自報
MMBench-Video2.0%自報

Reasoning

Hallusion Bench55.2%自報

AA 評測指數

暫無 AA 評測資料

LLM Stats 分類評分

Structured Output
80
Text-to-image
80
Image To Text
80
Spatial Reasoning
70
Grounding
70
Healthcare
70
Reasoning
70
Vision
60
Long Context
60
Math
60
Multimodal
60
Video
50
General
50
Agents
40

定價

輸入價格$0.25 / 1M tokens
輸出價格$0.75 / 1M tokens
混合價格(3:1)$0.375 / 1M tokens

速度

暫無速度資料

可用提供商

(LS 內部計價單位)

暫無提供商資料

外部連結