Which model leads OCRBench v2_en?

Q: Which model leads OCRBench v2_en?

Qwen3 VL 32B Reasoning tops the board at 68.4

OCRBench v2_en Leaderboard — Model Scores

Pro Mode — raw scores, one benchmark under the lensofficial card harness

The field · 27 measured

33.968.4

27 models measured, most on the official card harness. Qwen3 VL 32B Reasoning tops the board at 68.4.

27 measured·Lifecycle Updated Oct 8, 2026Score basis

#◆Vendor$/M in$/M out

Leader — ignites whiteField — fill cools with scoreIndependently verified only0 of 27Bars show 33.9–68.4 · ◆ solid = first-party or better · outlined = arm's-length

01Qwen3 VL 32B ReasoningAlibaba68.4$0.16/M$0.64/M$0.006

02Qwen3 VL 32B InstructAlibaba67.4$0.16/M$0.64/M$0.006

03Qwen3 VL 235B A22B InstructAlibaba67.1$0.40/M$1.6/M$0.015

04Nemotron 3 Nano OmniNVIDIA67.043P$0.30/M$0.90/M$0.009

05Qwen3 VL 235B A22B ReasoningAlibaba66.8$0.40/M$4.0/M$0.033

06Qwen3 VL 8B InstructAlibaba65.4$0.18/M$0.70/M$0.007

07Qwen3 VL Thinking (8B)Alibaba63.9$0.18/M$2.1/M$0.018

08Qwen3 VL 4B InstructAlibaba63.7———

09Qwen3 VL 30B A3B InstructAlibaba63.2$0.20/M$0.80/M$0.008

10Qwen3 VL 30B A3B ReasoningAlibaba62.6$0.20/M$2.4/M$0.021

11Qwen3 VL 4B (Reasoning)Alibaba61.8———

12Qwen2.5 VL 72BAlibaba61.5———

13Qwen3.5 4BAlibaba58.7$0.03/M$0.15/M$0.002

14Qwen2.5 VL 32B InstructAlibaba57.2———

15NVIDIA Nemotron Nano 12B v2 VLNVIDIA54.83P$0.20/M$0.60/M$0.007

16InternVL3_5-4BOpenGVLab49.1———

17Gemma 4 E4BGoogle48.8$0.02/M$0.10/M$0.001

18Qwen3.5 2B+2 altsAlibaba48.1———

19LFM2.5-VL-3B+1 altLiquid AI47.5———

20InternVL3_5-2BOpenGVLab45.5———

21LFM2-VL-3BLiquid AI43.9———

22Gemma 4 E2B+2 altsGoogle43.5———

23Qwen3 VL 2B Instruct+1 altAlibaba41.7———

24LFM2.5-VL-1.6B+1 altLiquid AI41.5———

25Ministral 3 3B+1 altMistral41.4$0.10/M$0.10/M$0.002

26North-Micro-Vision-Instruct+2 altsCohere36.7———

27Phi 3.5 Vision Instruct+1 altMicrosoft33.9———

◆ solid = independently verified · aggregator attested · vendor attributed — outlined = cross-reference · source-attributed. Dimmed ◆ = confidence inferred. “+N alts” = alternate disclosures in the drawer with why-demoted. Prices per million tokens; $/point = input price ÷ score. Every row opens its model dossier.

Verification: 27 models scored · 0 independently verified · 27 vendor-reported · 0 with source disagreement. How these tiers are assigned