How many tracked prints does the MMBench family have?

MMBench leaderboard — the family board.

12 TRACKED VARIANTS · 1 FIRST-CLASS BOARD

One family, 12 tracked prints. Scoring configs and splits of the base board render as sections of this page, each under its own honest label.

● LIVE DATA15 MODELS ON THIS BOARDLEADER STEP3 VL 10B AT 91.8

The board — MMBenchTOP 15 OF 15 MEASURED · BEST SCORE PER MODEL, DEFAULT VARIANT

#ModelVendorScoreVLast seen

01Step3 VL 10BStepFun91.82026-09-03

02Qwen2.5 VL 72BAlibaba882026-09-03

03Qwen3.5 4BAlibaba87.12026-09-10

04Phi 4 Multimodal InstructMicrosoft86.72026-09-03

05Qwen2 VL 72B InstructAlibaba86.52026-09-03

06Phi 3.5 Vision InstructMicrosoft81.92026-09-03

07InternVL3_5-4BOpenGVLab81.12026-09-10

08LFM2.5-VL-3BLiquid AI812026-09-03

09Phi 3 Vision 128K InstructMicrosoft80.52026-10-18

10LFM2-VL-3BLiquid AI802026-09-10

11InternVL3_5-2BOpenGVLab76.22026-09-10

12Qwen3.5 2BAlibaba73.12026-09-10

13Gemma 4 E4BGoogle71.62026-09-10

14Gemma 4 E2BGoogle64.22026-09-10

15Claude 3 HaikuAnthropic62.42026-10-18

◆ IV — INDEPENDENTLY VERIFIED · ◆ AA — AGGREGATOR · ◆ VA — VENDOR · ◇ SA — SOURCE-ATTRIBUTED

Variants of this boardSCORING CONFIGS & SPLITS — NOT CITED STANDALONE, SO THEY LIVE HERE

MMBench (dev en)

VARIANT12 models · best score per model

#ModelVendorScoreVLast seen

01Gemini 1.5 ProGoogle87.92026-08-24

02InternVL-2-8BOpenGVLab872026-08-24

03GPT-4o miniOpenAI83.82026-08-24

04InternVL-2-4BOpenGVLab83.42026-08-24

05Claude 3.5 SonnetAnthropic82.32026-08-24

06Phi 3.5 Vision InstructMicrosoft81.92026-08-24

07InternVL3-2BOpenGVLab81.12026-08-09

08Qwen2.5 VL 3BAlibaba80.42026-08-09

09LFM2-VL-3BLiquid AI79.82026-08-09

10InternVL3_5-2BOpenGVLab78.22026-08-09

11LFM2.5-VL-450M-ExtractLiquid AI60.92026-08-09

12LFM2-VL-450MLiquid AI56.32026-08-09

MMBench (EN)

VARIANT9 models · best score per model

#ModelVendorScoreVLast seen

01Gemini 2.5 ProGoogle93.22026-06-05

02GLM-4.6V (106B-A12B)Z.ai92.82026-06-05

03Qwen3 VL 235B A22B ReasoningAlibaba92.72026-06-05

04Step3 VL 10BStepFun92.42026-06-05

05Seed 1.5 VLByteDance92.12026-06-05

06GLM-4.6V-Flash (9B)Z.ai91.02026-06-05

07Qwen3 VL Thinking (8B)Alibaba90.52026-06-05

08MiMo VL RL 2508 (7B)Xiaomi89.92026-06-05

09InternVL-3.5 (8B)OpenGVLab88.22026-06-05

MMBench DEV_EN_V11

VARIANT7 models · best score per model

#ModelVendorScoreVLast seen

01Qwen3.5 2BAlibaba762026-08-12

02Qwen3 VL 2B InstructAlibaba74.42026-08-12

03Phi 3.5 Vision InstructMicrosoft73.12026-08-12

04LFM2.5-VL-1.6BLiquid AI69.62026-08-12

05Gemma 4 E2BGoogle69.32026-08-12

06Ministral 3 3BMistral69.22026-08-12

07North-Micro-Vision-InstructCohere68.72026-08-13

MMBench EN-DEV-v1.1

VARIANT7 models · best score per model

#ModelVendorScoreVLast seen

01Qwen3.6 35B A3BAlibaba92.82026-08-24

02Qwen3.5 27BAlibaba92.62026-08-24

03Qwen3.6 27BAlibaba92.32026-08-24

04Qwen3.5 35B A3BAlibaba91.52026-08-24

05Gemma 4 31BGoogle90.92026-08-24

06Gemma 4 26B A4BGoogle892026-08-24

07Claude Sonnet 4.5Anthropic88.32026-08-24

MMBench-EN-v1.1 (Acc)

VARIANT7 models · best score per model

#ModelVendorScoreVLast seen

01Qwen2.5 VL 72BAlibaba88.32026-06-05

02Kimi VL A3B ThinkingMoonshot84.42026-06-12

03GPT-4oOpenAI83.12026-06-05

04Kimi VL A3B InstructMoonshot83.12026-06-12

05Gemma 3 27BGoogle78.92026-06-05

06GPT-4o miniOpenAI77.12026-06-05

07Gemma 3 12BGoogle74.62026-06-05

Other tracked prints

THIN BOARDSunder 5 published models each — listed, never hidden

PrintLeaderScoreModels

MMBench-CN-devMage-VL-4B82.04

MMBench-EN-devPhi 4 R V 15B84.24

MMBench-CN_testQwen2 VL 72B Instruct86.63

MMBench-EN_testQwen2 VL 72B Instruct86.53

MMBench-V1.1_testQwen2 VL 72B Instruct85.93

MMBench-VideoQwen2.5 VL 72B2.03

THE FAMILY BOARD · EVERY PRINT LABELED · EVERY SCORE SOURCE-ATTRIBUTED ON ITS OWN LIVE BOARDAll leaderboards The frontier board Models by vendor