多模態大模型排行榜:視覺大模型排名

依真人匿名對比投票的 Arena 得分排名(LMArena 視覺競技場):使用者上傳圖片並提問,兩個匿名模型作答後投票。

圖表理解第一資料 2026年10月19日
Claude Fable 5 (High)Anthropic · 1357

依 LMArena 視覺競技場 2026年10月19日 發布的資料,Anthropic 的 Claude Fable 5 (High) 以 1325 分(14,714 票)在 167 個模型中排第一。

完整榜單 · 圖表理解

共 123 個

預設依 LMArena 得分降序;± 為 95% 信賴區間半寬;分項欄為該模型在對應分榜的得分,空白為 –。

# 模型 LMArena · 圖表理解 授權
得分 名次 票數
1 Claude Fable 5 (High)Anthropic 1357±10 1 3,956 Proprietary
2 Claude Opus 5 (High)Anthropic 1348±10 2 4,520 Proprietary
3 Claude Opus 5.5 (High)Anthropic 1345±23 3 711 Proprietary
4 Gemini 3.8 Flash (High)Google 1345±18 4 1,159 Proprietary
5 Claude Fable 5.1 (Max)Anthropic 1343±18 5 1,179 Proprietary
6 Claude Opus 4.6Anthropic 1340±9 6 7,249 Proprietary
7 Claude Opus 4.7 (High)Anthropic 1336±9 7 6,240 Proprietary
8 Claude Opus 4.7Anthropic 1336±9 8 6,411 Proprietary
9 Claude Opus 4.6 (High)Anthropic 1332±9 9 5,987 Proprietary
10 Muse Spark 1.3 (Max)Meta 1330±17 10 1,352 Proprietary
11 Gemini 3.7 Flash (High)Google 1329±17 11 1,364 Proprietary
12 Claude Sonnet 5.5 (xHigh)Anthropic 1329±24 12 625 Proprietary
13 GPT-5.4 (High)OpenAI 1327±8 13 8,230 Proprietary
14 GPT-5.5OpenAI 1325±8 14 7,798 Proprietary
15 Gemini 3.5 Flash (Medium)Google 1323±11 15 3,618 Proprietary
16 Qwen3.8-MaxAlibaba 1323±12 16 2,497 Proprietary
17 Gemini 3.5 Flash (High)Google 1318±11 17 3,586 Proprietary
18 Muse Spark 1.3 (xHigh)Meta 1318±17 18 1,236 Proprietary
19 GPT-5.5 (High)OpenAI 1317±8 19 7,220 Proprietary
20 GLM-5.3-FlashZ.ai開放權重 1317±10 20 3,995 MIT
21 Muse Spark 1.2 (xHigh)Meta 1317±26 21 544 Proprietary
22 GPT-5.4OpenAI 1316±9 22 6,221 Proprietary
23 Claude Opus 4.8 (High)Anthropic 1315±10 23 4,883 Proprietary
24 Gemini 3.6 Flash (High)Google 1314±13 24 2,285 Proprietary
25 Gemini 3.1 Pro PreviewGoogle 1311±7 25 12,416 Proprietary
26 Muse SparkMeta 1309±15 26 1,600 Proprietary
27 SStep 5 PreviewStepFun 1308±29 27 423 Proprietary
28 Claude Sonnet 4.6Anthropic 1306±8 28 7,630 Proprietary
29 GPT-6.1 Sol (Max)OpenAI 1306±25 29 573 Proprietary
30 GPT-5.6 Sol (xHigh)OpenAI 1305±12 30 2,880 Proprietary
31 Muse Spark 1.1Meta 1305±11 31 3,340 Proprietary
32 Grok 4.5xAI 1304±11 32 3,227 Proprietary
33 Claude Opus 4.8Anthropic 1302±9 33 4,921 Proprietary
34 Gemini 3 ProGoogle 1301±12 34 2,978 Proprietary
35 Kimi K2.6Moonshot AI開放權重 1298±10 35 4,304 Modified MIT
36 GPT-6 Astra (Max)OpenAI 1297±19 36 966 Proprietary
37 DeepSeek-V4.1-Flash-20260910 (Max)DeepSeek開放權重 1297±19 37 1,085 MIT
38 GPT-5.6 Terra (xHigh)OpenAI 1295±11 38 3,090 Proprietary
39 Claude Sonnet 5 (High)Anthropic 1293±10 39 3,718 Proprietary
40 Qwen3.7-PlusAlibaba 1292±10 40 3,752 Proprietary
41 Gemini 3 FlashGoogle 1290±7 41 9,994 Proprietary
42 Dola Seed 2.0 ProByteDance 1289±12 42 2,846 Proprietary
43 Mistral Large 4Mistral AI 1288±37 43 243 Proprietary
44 GLM-5V-TurboZ.ai 1286±7 44 11,284 Proprietary
45 Grok 4.6 (High)xAI 1285±14 45 1,838 Proprietary
46 Qwen3.8-27BAlibaba開放權重 1285±14 46 1,863 Apache 2.0
47 GPT-5.2 Chat (20260210)OpenAI 1285±10 47 4,187 Proprietary
48 Gemini 3.5 Flash-LiteGoogle 1284±13 48 2,334 Proprietary
49 Gemma 4 31BGoogle開放權重 1283±8 49 10,238 Apache 2.0
50 Kimi K2.5 (Thinking)Moonshot AI開放權重 1282±8 50 7,385 Modified MIT
51 MiMo-V2.6-ProXiaomi開放權重 1281±27 51 473 MIT
52 Qwen3.5-397B-A17BAlibaba開放權重 1277±7 52 8,623 Apache 2.0
53 Gemini 3 Flash (Minimal)Google 1276±7 53 9,717 Proprietary
54 MiMo-V2.6-FlashXiaomi開放權重 1276±20 54 919 MIT
55 Gemini 2.5 ProGoogle 1274±7 55 9,768 Proprietary
56 GPT-5.5 InstantOpenAI 1273±14 56 2,193 Proprietary
57 GPT-5.6 Luna (xHigh)OpenAI 1273±12 57 2,789 Proprietary
58 GPT-6 Sol (Max)OpenAI 1270±21 58 844 Proprietary
59 MiMo-V2.5Xiaomi開放權重 1269±8 59 7,723 MIT
60 Qwen-VL-Max (2025-08-13)Alibaba 1268±31 60 345 Proprietary
61 Grok 4.20 Beta (Reasoning, 0309)xAI 1267±9 61 7,150 Proprietary
62 MiniMax-M3MiniMax開放權重 1266±9 62 5,289 MiniMax Community License
63 GPT-5.4 mini (High)OpenAI 1266±9 63 6,940 Proprietary
64 GPT-5.2 (High)OpenAI 1265±9 64 4,773 Proprietary
65 GPT-5.1 (High)OpenAI 1264±13 65 2,263 Proprietary
66 Gemma 4 26B A4BGoogle開放權重 1264±9 66 6,239 Apache 2.0
67 Grok 4.20 Multi-Agent Beta (0309)xAI 1261±9 67 6,514 Proprietary
68 Kimi K2.5 (Instant)Moonshot AI開放權重 1257±19 68 955 Modified MIT
69 MiMo-V2-OmniXiaomi 1256±12 69 2,813 Proprietary
70 Qwen3-VL-235B-A22B-InstructAlibaba開放權重 1255±11 70 2,727 Apache 2.0
71 ChatGPT-4o Latest (20250326)OpenAI 1255±10 71 3,746 Proprietary
72 Grok 4.7 (xHigh)xAI 1254±20 72 938 Proprietary
73 Mistral Medium 3.5Mistral AI開放權重 1254±16 73 1,538 Modified MIT
74 GPT-5 ChatOpenAI 1252±13 74 2,522 Proprietary
75 TInkling-SmallThinking Machines開放權重 1251±14 75 2,071 Apache 2.0
76 Gemini 2.5 Flash Preview (09-2025)Google 1250±19 76 963 Proprietary
77 Qwen3.5-122B-A10BAlibaba開放權重 1249±11 77 3,455 Apache 2.0
78 Qwen3.5-27BAlibaba開放權重 1249±8 78 7,927 Apache 2.0
79 GPT-5.2OpenAI 1246±9 79 5,108 Proprietary
80 GPT-6 Luna (Max)OpenAI 1245±20 80 905 Proprietary
81 Gemini 2.5 FlashGoogle 1245±7 81 8,547 Proprietary
82 GPT-5.1OpenAI 1244±12 82 2,467 Proprietary
83 Gemini 3.1 Flash-Lite PreviewGoogle 1242±8 83 7,269 Proprietary
84 Grok 4.3xAI 1241±8 84 7,408 Proprietary
85 Mistral Large 3Mistral AI開放權重 1240±13 85 2,334 Apache 2.0
86 ERNIE-5.0-Preview (1220)Baidu 1236±21 86 709 Proprietary
87 GPT-5 (High)OpenAI 1227±12 87 2,969 Proprietary
88 NNVIDIA-Nemotron-3.5-Super-VL-120B-A12B (NVFP4)NVIDIA開放權重 1223±24 88 651 OpenMDW-1.1
89 o3 (2025-04-16)OpenAI 1220±12 89 3,574 Proprietary
90 GPT-4.1 (2025-04-14)OpenAI 1215±13 90 2,758 Proprietary
91 Qwen3-VL-235B-A22B-ThinkingAlibaba開放權重 1214±27 91 464 Apache 2.0
92 GPT-5.4 nano (High)OpenAI 1210±9 92 6,840 Proprietary
93 Grok 4.1 Fast (Reasoning)xAI 1210±11 93 3,459 Proprietary
94 Grok 4 (0709)xAI 1204±13 94 2,440 Proprietary
95 GPT-5 mini (High)OpenAI 1202±15 95 1,936 Proprietary
96 Claude Sonnet 4 (Thinking 32K, 20250514)Anthropic 1200±37 96 228 Proprietary
97 o4-mini (2025-04-16)OpenAI 1197±13 97 2,826 Proprietary
98 Gemini 2.5 Flash-Lite Preview (Non-Thinking, 09-2025)Google 1197±19 98 984 Proprietary
99 Hunyuan Vision 1.5 ThinkingTencent 1194±26 99 485 Proprietary
100 Gemini 2.5 Flash-Lite Preview (Thinking, 06-17)Google 1186±13 100 2,436 Proprietary
101 SStep-1o Turbo (202506)StepFun 1185±29 101 385 Proprietary
102 GLM-4.6VZ.ai開放權重 1182±23 102 578 MIT
103 GPT-4.1 mini (2025-04-14)OpenAI 1182±14 103 2,365 Proprietary
104 Claude Opus 4 (Thinking 16K, 20250514)Anthropic 1179±35 104 279 Proprietary
105 Claude Sonnet 4 (20250514)Anthropic 1178±29 105 391 Proprietary
106 Claude Opus 4 (20250514)Anthropic 1177±26 106 461 Proprietary
107 Claude 3.7 Sonnet (Thinking 32K, 20250219)Anthropic 1176±33 107 301 Proprietary
108 Mistral Medium 3.1 (2508)Mistral AI 1175±11 108 3,741 Proprietary
109 Gemma 3 27B ITGoogle開放權重 1156±15 109 1,715 Gemma
110 SStep-3StepFun開放權重 1155±30 110 347 Apache 2.0
111 Mistral Medium 3 (2505)Mistral AI 1155±16 111 1,423 Proprietary
112 Claude 3.7 Sonnet (20250219)Anthropic 1153±32 112 291 Proprietary
113 GLM-4.5VZ.ai開放權重 1151±30 113 369 MIT
114 Gemini 2.0 Flash (001)Google 1149±18 114 1,109 Proprietary
115 GPT-5 nano (High)OpenAI 1141±26 115 518 Proprietary
116 Hunyuan Large VisionTencent 1135±35 116 257 Proprietary
117 Claude 3.5 Sonnet (20241022)Anthropic 1134±32 117 314 Proprietary
118 Llama 4 Maverick 17B 128E InstructMeta開放權重 1134±19 118 947 Llama 4
119 Mistral Small 3.2 (2506)Mistral AI開放權重 1125±18 119 1,166 Apache 2.0
120 Llama 4 Scout 17B 16E InstructMeta開放權重 1125±20 120 845 Llama
121 Mistral Small 3.1 24B Instruct (2503)Mistral AI開放權重 1119±16 121 1,627 Apache 2.0
122 Claude 3.5 Haiku (20241022)Anthropic 1077±33 122 319 Proprietary
123 Molmo 2 8BAi2開放權重 1072±39 123 284 Apache 2.0

得分資料:LMArena leaderboard-dataset(lmarena-ai,Hugging Face,CC-BY-4.0,vision 設定),圖表理解發布於 2026年10月19日。引用請註明來源。

資料來源與計算方法

  1. 得分為 Bradley-Terry 模型擬合的 Arena 分數,來自真人在同一圖片問題下對兩個匿名回答的投票。
  2. 分項(OCR、圖表、作業題、圖片描述、中文/英文等)是 LMArena 依提示詞類別單獨計算的分榜,不是綜合分的拆分,名次可能和綜合榜差很多。
  3. 沒有公開資料的欄位(價格、上下文、發布時間等)不顯示,空白為 –;不做推測。
  4. 每天自動同步一次;JSON 介面: /api/public/vision-leaderboard.php

常見問題

目前最強的多模態(視覺)大模型是哪個?

依 LMArena 視覺競技場 2026年10月19日 發布的資料,Anthropic 的 Claude Fable 5 (High) 以 1325 分(14,714 票)在 167 個模型中排第一。

各家廠商最好的視覺模型分別是哪個?

綜合榜各廠商最高:Anthropic Claude Fable 5 (High) (1325, #1)、Google Gemini 3.8 Flash (High) (1315, #7)、Alibaba Qwen3.8-Max (1313, #10)、Meta Muse Spark 1.3 (Max) (1308, #12)、OpenAI GPT-5.4 (High) (1303, #18)。

哪個模型 OCR 文字辨識最好?

LMArena OCR 分榜(123 個模型)第一是 Claude Fable 5 (High),1337 分。

最好的開源/開放權重視覺模型是哪個?

榜上有 50 個開放權重模型,最高的是 GLM-5.3-Flash(1294 分,授權 MIT)。

這個分數和 MMMU 等基準測試有什麼區別?

MMMU 等是固定題庫的自動評分;這裡是真人用自己的圖片和問題做盲測投票,反映實際使用中的偏好,兩者不能直接換算。