Model

Qwen-VL

OPEN

Alibaba · Qwen-VL family · released Aug 24, 2023

First Qwen VLM (ViT-bigG + Qwen-7B, 9.6B total); accepts image/text/bounding-box input and can output bounding boxes for grounding.

ReasoningCodingVisionFunction callingTool useAgentic

343.4

Elo · unrated

Parameters

9.6B

Active params

9.6B (dense)

Context

8K tokens

Architecture

OpenCLIP ViT-bigG vision encoder + Qwen-7B decoder, single-layer cross-attention position-aware adapter

License

Tongyi Qianwen LICENSE AGREEMENT (custom, free for research and commercial use)

Languages

—

API price (in/out)

No hosted API

Modalities

text · vision

Benchmark results

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

AI2DVision62.3%#40

best: Molmo 72B · 96.3%

ChartQAVision65.7%#39

best: MiniMax-VL-01 · 91.7%

DocVQAVision65.1%#45

best: Qwen2-VL-72B · 96.5%

Flickr30kVision85.8#2

best: CogVLM-17B · 94.9

GQAVision59.3%#2

best: QVQ-72B-Preview · 63.9%

OK-VQAVision58.6%#2

best: CogVLM-17B · 64.8%

RefCOCO+Vision83.1%#6

best: DeepSeek-VL2 · 91.2%

RefCOCOVision89.4%#11

best: Qwen3.5-Omni-Plus · 95.0%

SEED-BenchVision56.3%#6

best: NAVER HyperCLOVA X SEED Think 32B · 77.9%

TextVQAVision63.8%#26

best: Molmo 2 8B · 85.7%

VQAv2Vision78.8%#9

best: Reka Edge (2026) · 88.4%

Run it locally

VRAM @ Q4

6 GB

VRAM @ FP16

19 GB

Fits on (Q4)

RTX 3060 12GBRTX 4070 Ti 16GBRTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

Throughput data unavailable.

Quantizations

GPTQ Int4

Qwen-VL family

Elo progression across releases

Qwen-VLAug 2023343.4

Resources

HFQwen/Qwen-VL GHQwenLM/Qwen-VL

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