Key takeaways

• Qwen3.5 Plus is designed as a large-scale multimodal language model for general-purpose AI tasks, processing text, image, and video input to produce text output. It excels in tasks requiring deep integration of vision, language, and code.
• The model is built on the Qwen3-Next architecture, incorporating a hybrid attention mechanism (Gated DeltaNet and Gated Attention) and a sparse Mixture of Experts (MoE) design, featuring native multimodal processing through early text-visual fusion technology.
• Notable capabilities include processing video content up to 2 hours long and strong performance across various mathematical problem-solving domains and diverse academic disciplines, achieving high accuracy in literature, marketing, sociology, math, history, and finance. However, it struggles with highly complex combinatorial problems, detailed NMR/IR interpretation in organic chemistry, and numerical precision in technical calculations.
• On the MATH-500 benchmark, Qwen3.5 Plus achieved an accuracy of 97.6%, demonstrating strong performance in fundamental algebraic manipulation, geometry, number theory, and precalculus.
• The model struggled on the Multimodal Understanding benchmark, achieving an accuracy of 81.11%. Specifically, it showed weaknesses in Art (54.55% accuracy) and Agriculture (55.56% accuracy), as well as occasional numerical calculation errors in engineering domains and misinterpretation of visual details.

Technical specifications

Context length

1M tokens

Max output

—

Architecture

MoE

Modalities

text+image+video->text

Parameters

~397B

Open weights

Yes

Licence

Apache 2.0

Regions

China

Vendor

Qwen

Provider model key

qwen/qwen3.5-plus-2026-04-20