跳轉到主要內容

Qwen3 VL 235B A22B (Reasoning)

AlibabaQwenOpen WeightApache 2.0 · Commercial OK

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.

發布日期
2025-09-23
參數規模
236.0B
上下文長度
131K
支援模態
image, text, video

能力雷達圖

42
general
38
coding
86
reasoning
51
science估算
70
agents
100
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24
66.0
LS
代码能力榜171
47.0
AA
通用能力榜146
59.0
AA
数学推理49
89.0
AA
多模态榜64
67.0
LS
推理能力37
75.0
LS
科学能力137
56.0
AA

基準測試分數 (LLM Stats)

3d

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

Agents

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

Chemistry

SuperGPQA64.3%自報

Code

Design2Code0.93 / 100自報

Communication

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

Creativity

Creative Writing v385.7%自報

Embodied

EmbSpatialBench0.84 / 100自報
RoboSpatialHome0.74 / 100自報

Factuality

SimpleQA44.4%自報

Finance

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

General

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

Grounding

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

Healthcare

VideoMMMU80.0%自報

Image To Text

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

Instruction Following

MIABench0.93 / 100自報

Language

CharadesSTA63.5%自報

Long Context

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

Math

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

Multimodal

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

Reasoning

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

Spatial Reasoning

RealWorldQA81.3%自報

Vision

ODinW43.2%自報

AA 評測指數

Math Index
88.3
Intelligence Index
27.6
Coding Index
20.9
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 分類評分

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

定價

輸入價格$0.84 / 1M tokens
輸出價格$6.175 / 1M tokens
混合價格(3:1)$2.174 / 1M tokens

速度

Tokens/秒30.0 tokens/s
首Token延遲1.35s
首回答延遲68.10s

可用提供商

(LS 內部計價單位)
提供商輸入價格輸出價格
DeepInfra450K3.5M
Novita980K4.0M

外部連結