Atlas

Models

← All models

Qwen3 VL 30B A3B Instruct

Alibaba · 2025-10-04 · 31.1B parameters

Qwen3-VL-30B-A3B-Instruct is Alibaba's non-thinking instruction-tuned 30B-A3B mixture-of-experts vision-language model. It accepts text, images, and video and supports multilingual OCR, spatial grounding, and visual-agent interaction. Its native context is 262,144 tokens and can be extended to 1,000,000 tokens with YaRN.

Benchmark scores

BenchmarkScore
AA-LCR23.7
AA-Omniscience Index-63.0
AIME 202572.3
Artificial Analysis Intelligence Index10.0
Artificial Analysis Omniscience Accuracy15.5
Artificial Analysis Omniscience Hallucination Rate92.9
Artificial Analysis Openness Index50.0
BuseyBench SVG0.4
Capability133.0
CritPt0.0
GPQA Diamond69.5
Humanity's Last Exam (Text-Only)6.4
IFBench (Artificial Analysis)33.1
LiveCodeBench (Artificial Analysis)47.6
MMLU Pro76.4
MMMU Pro62.1
ObviousBench46.5
Phare Average Safety59.3
Phare Bias Resistance53.6
Phare Debunking89.1
Phare Debunking (English)89.9
Phare Debunking (French)88.1
Phare Debunking (Spanish)89.4
Phare Encoding Jailbreaks33.6
Phare Encoding Jailbreaks (English)34.6
Phare Encoding Jailbreaks (French)32.7
Phare Encoding Jailbreaks (Spanish)33.6
Phare Factuality44.9
Phare Factuality (English)60.7
Phare Factuality (French)41.5
Loading Atlas data…