Atlas

Models

← All models

Qwen3 VL 4B Thinking

Alibaba · 2025-10-11 · 4.4B parameters

Qwen3-VL-4B-Thinking is the reasoning-enabled variant of Alibaba's compact 4B Qwen3-VL vision-language model. It targets lower-resource multimodal reasoning and visual-agent tasks while retaining OCR, grounding, and image/video understanding capabilities. Its context is 262,144 tokens natively and expandable to 1 million tokens.

Benchmark scores

BenchmarkScore
AA-LCR21.3
AA-Omniscience Index-69.3
AIME 202525.7
Artificial Analysis Intelligence Index7.9
Artificial Analysis Omniscience Accuracy12.2
Artificial Analysis Omniscience Hallucination Rate92.8
Artificial Analysis Openness Index50.0
Capability126.6
CritPt0.0
GPQA Diamond49.4
Humanity's Last Exam (Text-Only)4.4
IFBench (Artificial Analysis)36.6
LiveCodeBench (Artificial Analysis)32.0
MMLU Pro70.0
MMMU Pro52.0
SciCode (Artificial Analysis)17.1
Terminal-Bench Hard1.5
τ²-bench (Telecom)15.5
Loading Atlas data…