Atlas

Models

← All models

Qwen3 VL 4B Instruct

Alibaba · 2025-10-15 · 4.4B parameters

Qwen3-VL-4B-Instruct is the instruction-tuned, dense 4B vision-language model in Alibaba's Qwen3-VL family. This compact variant carries the family's image/video understanding, multilingual OCR, grounding, and visual-agent capabilities. Its context is 262,144 tokens natively and expandable to 1 million tokens.

Benchmark scores

BenchmarkScore
AA-LCR13.0
AA-Omniscience Index-75.8
AIME 202537.0
Artificial Analysis Intelligence Index4.1
Artificial Analysis Omniscience Accuracy10.3
Artificial Analysis Omniscience Hallucination Rate96.1
Artificial Analysis Openness Index50.0
Capability121.7
CritPt0.0
GPQA Diamond37.1
Humanity's Last Exam (Text-Only)3.7
IFBench (Artificial Analysis)31.8
LiveCodeBench (Artificial Analysis)29.0
MMLU Pro63.4
MMMU Pro43.9
SciCode (Artificial Analysis)13.7
Terminal-Bench Hard0.0
τ²-bench (Telecom)23.4
Loading Atlas data…