Vision Model Leaderboard: Best Vision-Language Models

Ranked by LMArena Vision Arena scores: users upload an image, ask a question, and vote between two anonymous models’ answers.

Overall OCR Diagrams Homework Captioning Entity recognition Creative writing Humor English Chinese
#1 OverallData Oct 19, 2026
Claude Fable 5 (High)Anthropic · 1325

In LMArena Vision Arena data published Oct 19, 2026, Anthropic’s Claude Fable 5 (High) ranks #1 of 167 models with 1325 (14,714 votes).

Full leaderboard · Overall

50 models
All Open weights

Sorted by LMArena score by default; ± is half the 95% confidence interval. Category columns show the model’s score on that sub-board; – means no score.

# Model LMArena · Overall Category scores License
Score Rank Votes OCR Diagrams Homework Captioning
1 MPhi-3 Vision 128K InstructMicrosoftOpen 812±19 167 1,401 – – – – MIT
2 MPhi-3.5 Vision InstructMicrosoftOpen 851±16 166 2,592 – – – – MIT
3 OMiniCPM-V 2.6OpenBMBOpen 912±16 165 1,987 – – – – Apache 2.0
4 LLLaVA-OneVision Qwen2 72B OVOpen 923±18 164 1,321 – – – – Apache 2.0
5 CogVLM2 Llama3 Chat 19BZ.aiOpen 928±15 163 1,991 – – – – CogVLM2
6 LLLaVA 1.6 34BOpen 936±12 162 4,531 – – – – Apache 2.0
7 Nnvila-internal-15b-v1NVIDIAOpen 940±20 161 1,077 – – – – -
8 OInternVL2 4BOpenGVLabOpen 941±13 160 3,703 – – – – MIT
9 Molmo 7B-D (0924)Ai2Open 959±14 158 2,815 – – – – Apache 2.0
10 Llama 3.2 11B Vision InstructMetaOpen 966±11 157 4,817 – – – – Llama 3.2
11 Qwen2-VL-7B-InstructAlibabaOpen 990±11 153 5,766 – – – – Apache 2.0
12 Aya Vision 32BCohereOpen 996±22 151 847 – – – – CC-BY-NC-4.0
13 Llama 3.2 90B Vision InstructMetaOpen 1000±9 149 8,682 – – – – Llama 3.2
14 OInternVL2 26BOpenGVLabOpen 1005±13 148 5,148 – – – – MIT
15 Pixtral 12B (2409)Mistral AIOpen 1009±10 146 7,511 – – – – Apache 2.0
16 Molmo 72B (0924)Ai2Open 1012±13 145 3,048 – – – – Apache 2.0
17 Qwen2-VL-72BAlibabaOpen 1047±10 141 5,937 – – – – Qwen
18 Molmo 2 8BAi2Open 1081±20 136 1,190 1058 1072 1069 – Apache 2.0
19 Pixtral Large (2411)Mistral AIOpen 1089±9 134 5,423 – – – – MRL
20 Qwen2.5-VL-72B-InstructAlibabaOpen 1107±11 130 3,768 – – – – Qwen
21 Llama 4 Scout 17B 16E InstructMetaOpen 1118±9 129 6,446 1117 1125 1148 – Llama
22 Mistral Small 3.1 24B Instruct (2503)Mistral AIOpen 1136±8 125 28,490 1132 1119 1158 1153 Apache 2.0
23 Mistral Small 3.2 (2506)Mistral AIOpen 1142±9 121 10,843 1134 1125 1173 1059 Apache 2.0
24 Llama 4 Maverick 17B 128E InstructMetaOpen 1142±9 122 6,923 1143 1134 1164 – Llama 4
25 Qwen2.5-VL-32B-InstructAlibabaOpen 1152±15 119 1,490 – – – – Apache 2.0
26 GLM-4.5VZ.aiOpen 1154±12 118 3,243 1151 1151 1163 – MIT
27 GLM-4.6VZ.aiOpen 1164±14 113 2,455 1171 1182 1155 – MIT
28 Gemma 3 27B ITGoogleOpen 1165±8 112 16,982 1156 1156 1160 1151 Gemma
29 SStep-3StepFunOpen 1176±12 106 3,193 1168 1155 1183 – Apache 2.0
30 NNVIDIA-Nemotron-3.5-Super-VL-120B-A12B (NVFP4)NVIDIAOpen 1205±13 92 2,509 1216 1223 1229 – OpenMDW-1.1
31 Qwen3-VL-235B-A22B-ThinkingAlibabaOpen 1207±12 91 2,387 1212 1214 1234 – Apache 2.0
32 Mistral Large 3Mistral AIOpen 1220±8 84 9,254 1228 1240 1216 – Apache 2.0
33 Mistral Medium 3.5Mistral AIOpen 1222±9 83 5,375 1233 1254 1230 – Modified MIT
34 TInkling-SmallThinking MachinesOpen 1237±8 75 8,074 1243 1251 1257 – Apache 2.0
35 Qwen3.5-27BAlibabaOpen 1241±5 72 30,272 1246 1249 1256 1170 Apache 2.0
36 Qwen3.5-122B-A10BAlibabaOpen 1244±7 70 13,236 1251 1249 1261 – Apache 2.0
37 MiMo-V2.5XiaomiOpen 1247±6 66 29,381 1256 1269 1262 – MIT
38 Qwen3-VL-235B-A22B-InstructAlibabaOpen 1247±7 67 12,724 1253 1255 1277 1281 Apache 2.0
39 MiniMax-M3MiniMaxOpen 1251±6 62 19,844 1259 1266 1268 – MiniMax Community License
40 Kimi K2.5 (Instant)Moonshot AIOpen 1254±11 60 4,111 1259 1257 1264 – Modified MIT
41 MiMo-V2.6-FlashXiaomiOpen 1256±11 59 3,618 1270 1276 1283 – MIT
42 Gemma 4 26B A4BGoogleOpen 1260±6 56 23,922 1268 1264 1293 1297 Apache 2.0
43 Qwen3.5-397B-A17BAlibabaOpen 1262±5 54 32,657 1271 1277 1292 1230 Apache 2.0
44 MiMo-V2.6-ProXiaomiOpen 1263±15 53 1,754 1270 1281 1264 – MIT
45 Qwen3.8-27BAlibabaOpen 1269±8 44 7,614 1281 1285 1295 – Apache 2.0
46 Kimi K2.5 (Thinking)Moonshot AIOpen 1269±6 46 28,249 1276 1282 1297 1257 Modified MIT
47 Gemma 4 31BGoogleOpen 1276±6 39 40,219 1282 1283 1310 1218 Apache 2.0
48 DeepSeek-V4.1-Flash-20260910 (Max)DeepSeekOpen 1278±10 38 4,100 1282 1297 1307 – MIT
49 Kimi K2.6Moonshot AIOpen 1282±7 34 16,368 1291 1298 1310 – Modified MIT
50 GLM-5.3-FlashZ.aiOpen 1294±6 24 16,254 1299 1317 1318 – MIT

Scores: LMArena leaderboard-dataset (lmarena-ai on Hugging Face, CC-BY-4.0, vision config), Overall published Oct 19, 2026. Please credit the source when citing.

Sources & methodology

  1. Scores are Bradley-Terry Arena ratings fitted from human votes between two anonymous answers to the same image prompt.
  2. Categories (OCR, diagrams, homework, captioning, Chinese/English…) are separate LMArena boards computed on prompts of that type, not a breakdown of the overall score; ranks can differ a lot.
  3. Fields without public data (price, context length, release date…) are not shown; blanks are –. Nothing is estimated.
  4. Synced daily. JSON API: /api/public/vision-leaderboard.php

This week

Collecting data: 1/7 days recorded; weekly changes appear after 7 days

Synced automatically once a day; ranks are compared only between real snapshots, never backfilled.

FAQ

What is the best vision-language model right now?

In LMArena Vision Arena data published Oct 19, 2026, Anthropic’s Claude Fable 5 (High) ranks #1 of 167 models with 1325 (14,714 votes).

What is each vendor’s best vision model?

Top overall entry per vendor: 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).

Which model is best at OCR?

On LMArena’s OCR board (123 models), Claude Fable 5 (High) is #1 with 1337.

What is the best open-weight vision model?

50 open-weight models are listed; the top one is GLM-5.3-Flash (1294, MIT).

How is this different from benchmarks like MMMU?

MMMU and similar benchmarks auto-score a fixed question set; this board comes from blind human votes on people’s own images and questions, reflecting real-use preference. They are not convertible.