Vision-language

Qwen2.5-VL-72B-Instruct

By Alibaba Group, in the Qwen-VL family.

Released
28 Jan 2025
Status
Legacy
Context
128K
Max output
—
Input per 1M tokens
—
Output per 1M tokens
—
Open weights
Yes
License
qwen
Parameters
72B
Knowledge cutoff
—
Reasoning
No
Modalities
text, image, video → text

Category scores

Out of 100, from independent and third-party results only, by method category-2026-09-29.1.

Category scores
CategoryScoreRankCoverage
Vision & multimodal understanding62.0Not ranked1 of 3
Human preference16.7Not ranked1 of 5

Benchmark results

Every result we hold for this model, with who ran it and where it was published.

Knowledge & reasoning

Knowledge & reasoning results
BenchmarkMetricValueProvenanceSettingEvaluatedSource
MMLU-ProAccuracy0.6%Self-reported——EvalEval via Hugging Face Community Evals

Computer & browser use

Computer & browser use results
BenchmarkMetricValueProvenanceSettingEvaluatedSource
OSWorld-Verified% tasks completed5.0%Third party——OS World Website via Epoch AI
OSWorld-Verified% tasks completed4.4%Third party——OS World Website via Epoch AI

Vision & multimodal understanding

Vision & multimodal understanding results
BenchmarkMetricValueProvenanceSettingEvaluatedSource
GeoBenchCountry accuracy62.0%Third party——GeoBench leaderboard via Epoch AI

Human preference

Human preference results
BenchmarkMetricValueProvenanceSettingEvaluatedSource
Vision Arena (arena.ai)Arena score (Elo scale)1121Independent—28 Sep 2026Arena
Vision Arena (arena.ai)Votes3768Independent—28 Sep 2026Arena

Other Qwen-VL versions

Sources