跳转到主要内容

Qwen3 VL 235B A22B (Reasoning)

AlibabaQwen开源权重Apache 2.0 · 商用许可

描述

Qwen3-VL-235B-A22B-Thinking is the most powerful vision-language model in the Qwen series, featuring 236B parameters with MoE architecture for reasoning-enhanced multimodal understanding. Key capabilities include: Visual Agent (operates PC/mobile GUIs, recognizes elements, invokes tools), Visual Coding (generates Draw.io/HTML/CSS/JS from images/videos), Advanced Spatial Perception (2D grounding and 3D grounding for spatial reasoning and embodied AI), Long Context & Video Understanding (native 256K context expandable to 1M, handles hours-long video with second-level indexing), Enhanced Multimodal Reasoning (excels in STEM/Math with causal analysis), Upgraded Visual Recognition (celebrities, anime, products, landmarks, flora/fauna), and Expanded OCR (32 languages, robust in low light/blur/tilt). Architecture innovations include Interleaved-MRoPE for positional embeddings, DeepStack for multi-level ViT feature fusion, and Text-Timestamp Alignment for precise video temporal modeling.

发布日期
2025-09-23
参数规模
236.0B
上下文长度
131K
支持模态
image, text, video

能力雷达图

38
general
59
coding
86
reasoning
51
science估算
70
agents
100
multimodal

Science 在缺少专门科学评测时使用推理能力代理估算。

排行榜排名

领域#排名分数来源
智能体能力模型榜19
66.0
LS
代码能力榜158
55.0
AA
通用能力榜165
55.0
AA
数学推理49
89.0
AA
多模态榜73
67.0
LS
推理能力40
75.0
LS
科学能力155
54.0
AA

基准测试分数 (LLM Stats)

3d

Objectron0.71 / 100自报
BLINK67.1%自报
ARKitScenes0.54 / 100自报
SUNRGBD0.35 / 100自报
Hypersim0.11 / 100自报

Agents

SIFO0.77 / 100自报
BFCL-v371.9%自报
SIFO-Multiturn0.71 / 100自报
OSWorld-G0.68 / 100自报
OSWorld38.1%自报

Chemistry

SuperGPQA64.3%自报

Code

Design2Code0.93 / 100自报

Communication

MM-MT-Bench8.50 / 100自报
WritingBench86.7%自报
Multi-IF79.1%自报

Creativity

Creative Writing v385.7%自报

Embodied

EmbSpatialBench0.84 / 100自报
RoboSpatialHome0.74 / 100自报

Factuality

SimpleQA44.4%自报

Finance

MMLU90.6%自报
MMLU-Pro83.8%自报
MMLU-ProX80.6%自报

General

MMLU-Redux93.7%自报
IFEval88.2%自报
MMMUval80.6%自报
Include80.0%自报
LiveBench 2024112579.6%自报
MMStar78.7%自报
LiveCodeBench v670.1%自报
MMMU-Pro69.3%自报
SimpleVQA0.61 / 100自报

Grounding

ScreenSpot95.4%自报
RefCOCO-avg0.92 / 100自报
RefSpatialBench0.70 / 100自报
ScreenSpot Pro61.8%自报

Healthcare

VideoMMMU80.0%自报

Image To Text

OCRBench87.5%自报
OCRBench-V2 (en)66.8%自报
OCRBench-V2 (zh)63.5%自报

Instruction Following

MIABench0.93 / 100自报

Language

CharadesSTA63.5%自报

Long Context

MLVU83.8%自报
LVBench63.6%自报
MMLongBench-Doc0.56 / 100自报

Math

AIME 202589.7%自报
MathVista-Mini85.8%自报
MathVerse-Mini0.85 / 100自报
HMMT2577.4%自报
MathVision74.6%自报
Humanity's Last Exam13.6%自报

Multimodal

DocVQAtest96.5%自报
MMBench-V1.190.6%自报
InfoVQAtest89.5%自报
AI2D89.2%自报
CC-OCR81.5%自报
MuirBench80.1%自报
VideoMME w/o sub.79.0%自报
CharXiv-R66.1%自报
VisuLogic0.34 / 100自报
ZEROBench-Sub0.28 / 100自报
ZEROBench0.04 / 100自报

Reasoning

ZebraLogic97.3%自报
CountBench0.94 / 100自报
Hallusion Bench66.7%自报
ERQA52.5%自报

Spatial Reasoning

RealWorldQA81.3%自报

Vision

ODinW43.2%自报

AA 评测指数

Math Index
88.3
Intelligence Index
20.6
Aime 25
0.9
Mmlu Pro
0.8
Gpqa
0.8
Livecodebench
0.6
Lcr
0.6
Ifbench
0.6
Tau2
0.5
Scicode
0.4
Terminalbench Hard
0.1
Hle
0.1

LLM Stats 分类评分

Communication
3
Multimodal
100
Creativity
90
Writing
90
Instruction Following
80
Language
80
Legal
80
Math
80
Structured Output
80
Embodied
80
Finance
80
Grounding
80
Healthcare
80
Text-to-image
80
Video
80
Image To Text
70
Long Context
70
Reasoning
70
Spatial Reasoning
70
General
70
Tool Calling
70
Vision
70
Physics
60
Agents
60
Chemistry
60
Economics
60
Factuality
40
3d
40

定价

输入价格$0.84 / 1M tokens
输出价格$6.175 / 1M tokens
混合价格(3:1)$2.174 / 1M tokens

速度

Tokens/秒57.2
首Token延迟1.16s
首回答延迟36.11s

供应商价格排行

供应商价格排行

10 个供应商

最便宜: Venice AI最贵: NovitaAI
供应商输入输出
1Venice AI最便宜
$0.25
$1.5
2OpenRouter
$0.26
$2.6
3Kilo Gateway
$0.26
$2.6
4Alibaba (China)
$0.28671
$1.14682
5SiliconFlow (China)
$0.45
$3.5
6SiliconFlow
$0.45
$3.5
7NanoGPT
$0.5
$6
8LLM Gateway
$0.5
$2
9Alibaba
$0.7
$2.8
10NovitaAI
$0.98
$3.95

比较该模型在不同 API 供应商之间的定价。

外部链接