Key takeaways

• Qwen3 VL 235B A22B Thinking is a multimodal Mixture of Experts (MoE) model with 175 billion parameters, designed for multimodal reasoning in STEM and math, unifying text generation and visual understanding across images and video.
• The model is optimized for robust perception, spatial understanding (2D/3D grounding), and long-form visual comprehension, featuring a context length of 131072 tokens and open-weights.
• The model demonstrates strong general knowledge across many academic subjects, particularly in pure sciences like abstract algebra and astronomy, with high accuracy in many factual recall questions.
• It struggles significantly with factual accuracy, multi-faceted reasoning, complex inference, temporal reasoning, and synthesizing information from multiple sources, leading to high hallucination rates and misinterpretation of question intent.
• On the Massive Multitask Language Understanding (MMLU) benchmark, it showed excellent performance in moral_scenarios (accuracy: 0.9458) and strong performance in clinical_knowledge (accuracy: 0.9090) and abstract algebra (accuracy: 0.9166), but on the SimpleQA benchmark, it achieved only 37.86% accuracy, often fabricating information and misinterpreting question intent.

Technical specifications

Context length

131k tokens

Max output

—

Architecture

MoE

Modalities

text+image->text

Parameters

~175B

Open weights

Yes

Licence

Open

Regions

China

Vendor

Qwen

Provider model key

qwen/qwen3-vl-235b-a22b-thinking