Acropolis

Qwen2.5 VL 72B Instruct

AlibabaGrade B. Evidence grade B of A to D: good.Open weightsReleased 19 Sept 2024Consistent across benchmarks
Compare with peers

Domain profile

Median within-board percentile by capability domain across boards with sufficient peer overlap.

  • Agentic (1 boards): 6 (scale 0 to 100)Agentic (1 boards)6
  • Vision (3 boards): 24 (scale 0 to 100)Vision (3 boards)24

Benchmark appearances

Best verified run per board, with confidence intervals and Pareto frontier status.

BenchmarkRankScoreCostStatusObserved
OSWorld

Epoch AI Benchmarking Hub · Default

#35 / 38
5%UnknownStandardNot stated
OSWorld

Epoch AI Benchmarking Hub · Default

#36 / 38
4.4%UnknownStandardNot stated
LMArena Vision

LMArena · Chinese · Default

#112 / 135
1,107[1,051 to 1,162]UnknownStandard9 Nov 2026
LMArena Vision (style control)

LMArena · Chinese · Default

#114 / 135
1,126[1,072 to 1,181]UnknownStandard9 Nov 2026
LMArena Vision

LMArena · Overall · Default

#130 / 167
1,107[1,096 to 1,118]UnknownStandard9 Nov 2026
LMArena Vision

LMArena · English · Default

#131 / 167
1,115[1,101 to 1,128]UnknownStandard9 Nov 2026
LMArena Vision (style control)

LMArena · Overall · Default

#131 / 167
1,121[1,110 to 1,132]UnknownStandard9 Nov 2026
LMArena Vision (style control)

LMArena · English · Default

#132 / 167
1,124[1,110 to 1,138]UnknownStandard9 Nov 2026

Closest head-to-head rivals

Direct pairwise win rates for Qwen2.5 VL 72B Instruct across shared benchmark boards.