Model

Qwen2-VL-72B

OPEN

Alibaba · Qwen2-VL family · released Sep 19, 2024

Gen-2 VLM flagship with dynamic-resolution ViT and M-ROPE for long video; benchmark scores below are for the -Instruct checkpoint.

ReasoningCodingVisionFunction callingTool useAgentic

931.3

Elo · rank #226

Parameters

73.4B

Active params

73.4B (dense)

Context

32K tokens

Architecture

Dense transformer + 675M-param ViT vision encoder (GQA, M-ROPE, dynamic resolution)

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

BIG-Bench HardReasoning69.2%#60

best: ERNIE 4.5 300B-A47B · 94.3%

ChartQAVision88.3%#10

best: MiniMax-VL-01 · 91.7%

DocVQAVision96.5%#1

best: this model · 96.5%

GPQA DiamondReasoning38.8%#225

best: GPT-5.6 · 94.6%

IFEvalReasoning54.9%#127

best: Gemma 4 26B A4B · 98.5%

MathVistaVision70.5%#26

best: Seed 2.1 Pro · 90.7%

MMBench (Chinese)Vision86.6%#5

best: ERNIE 4.5 VL 424B-A47B · 90.9%

MMBench (English)Vision86.5%#13

best: Qwen3.5-397B-A17B · 93.7%

MMLU-ProKnowledge57.2%#119

best: Claude Fable 5 · 91.5%

MMMUVision64.5%#67

best: Claude Fable 5 · 89.3%

OCRBenchVision877#5

best: InternVL3-78B · 906

RefCOCO+Vision90.1%#4

best: DeepSeek-VL2 · 91.2%

TextVQAVision85.5%#2

best: Molmo 2 8B · 85.7%

Video-MMEVision77.8%#12

best: Seed 2.1 Pro · 89.2%

Run it locally

VRAM @ Q4

48 GB

VRAM @ FP16

147 GB

Fits on (Q4)

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

~48GB Q4 weights exceed a single RTX 4090's 24GB VRAM; requires multi-GPU or CPU offload.

Quantizations

GGUF · AWQ · GPTQ Int4 · GPTQ Int8

Qwen2-VL family

Elo progression across releases

Qwen2-VL-72BSep 2024931.3

Resources

HFQwen/Qwen2-VL-72B GHQwenLM/Qwen2-VL DOCSqwenlm.github.io/blog/qwen2-vl

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