Model

Qwen2.5-VL-72B

OPEN

Alibaba · Qwen2.5-VL family · released Jan 28, 2025

Gen-2.5 vision flagship with agentic GUI/computer-use grounding and 1hr+ video understanding; 3B/7B/72B family.

ReasoningCodingVisionFunction callingTool useAgentic

1171.4

Elo · rank #198

Parameters

73B

Active params

73B (dense)

Context

32K tokens

Architecture

Dense Vision-Language Transformer (window-attention ViT + Qwen2.5 LLM, mRoPE)

License

Qwen (Tongyi Qianwen) License

Languages

—

API price (in/out)

No hosted API

Modalities

text · vision · video

Benchmark results

Bar shows position within the tracked field; marker = field best

AI2DVision88.4%#25

best: Molmo 72B · 96.3%

ChartQAVision89.5%#6

best: MiniMax-VL-01 · 91.7%

CharXivVision49.7%#37

best: Qwen3.8-Flash-Next · 90.6%

DocVQAVision96.4%#3

best: Qwen2-VL-72B · 96.5%

GPQA DiamondReasoning49.0%#200

best: GPT-6 Astra · 96.0%

GSM8KMath95.3%#7

best: Llama 3.1 405B · 96.8%

HumanEvalCoding87.8%#32

best: Claude Opus 4.5 · 99.4%

IFEvalReasoning86.3%#48

best: Gemma 4 26B A4B · 98.5%

MATH-500Math83.0%#62

best: GPT-5 · 99.4%

MathVisionVision38.1%#18

best: Seed 2.1 Pro · 92.6%

MathVistaVision74.8%#17

best: Seed 2.1 Pro · 90.7%

MMBench (Chinese)Vision87.9%#4

best: ERNIE 4.5 VL 424B-A47B · 90.9%

MMBench (English)Vision88.0%#10

best: Qwen3.5-397B-A17B · 93.7%

MMLU-ProKnowledge71.2%#89

best: Claude Fable 5 · 91.5%

MMLU-ReduxKnowledge85.9%#29

best: Qwen3.7-Max · 95.0%

MMMU-ProVision51.1%#50

best: Claude Opus 4.7 · 85.5%

MMMUVision70.2%#53

best: Claude Fable 5 · 89.3%

OCRBenchVision885#4

best: InternVL3-78B · 906

OSWorldAgents8.8%#5

best: Seed 2.1 Pro · 78.8%

RefCOCO+Vision88.9%#5

best: DeepSeek-VL2 · 91.2%

RefCOCOVision92.7%#6

best: Qwen3.5-Omni-Plus · 95.0%

TextVQAVision83.5%#6

best: Molmo 2 8B · 85.7%

Video-MMEVision79.1%#11

best: Seed 2.1 Pro · 89.2%

Run it locally

VRAM @ Q4

42 GB

VRAM @ FP16

144 GB

Fits on (Q4)

M3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

Throughput data unavailable.

Quantizations

GGUF · AWQ

Fine-tune it

Conditional / custom

QLoRA49.6 GB1× A100 80GB

LoRA155.5 GB1× B200 192GB

Full fine-tune1171.7 GB8× B200 192GB

QLoRA SFT on ~10k samples ≈ $51.34 (1× A100 80GB)

Plan a fine-tune →

Qwen2.5-VL family

Elo progression across releases

Qwen2.5-VL-72BJan 20251171.4

Resources

HFQwen/Qwen2.5-VL-72B-Instruct GHQwenLM/Qwen2.5-VL DOCSqwenlm.github.io/blog/qwen2.5-vl APIAvailable via Alibaba API

API price weights · each benchmark row carries its own source badge (see methodology)