Key takeaways

• Qwen3.6 Flash is a fast, efficient multimodal language model designed for agentic coding tasks, supporting text, image, and video input with a 1M token context window.
• It utilizes a novel hybrid MoE architecture combining Gated DeltaNet linear attention with standard Gated Attention, employing only 3B active parameters out of 35B total.
• The model demonstrates exceptional agentic coding performance, achieving 73.4% on SWE-bench Verified, and supports both thinking and non-thinking modes for complex reasoning.
• It shows strong mathematical problem-solving capabilities, with 95.9% accuracy on the MATH benchmark subset, excelling in standard algebraic manipulation, sequence/series, and number theory. It struggles with highly abstract combinatorial problems and nuanced interpretations in domains like Art Theory and Music.
• Performance is robust in tasks requiring direct information extraction from images or tables (e.g., scientific and engineering calculations), but challenges persist in complex multi-step algebraic simplification and precise interpretation of specialized domain contexts like medical diagrams or specific musical notation.

Technical specifications

Context length

1M tokens

Max output

—

Architecture

MoE

Modalities

text+image+video->text

Parameters

~35B

Open weights

Yes

Licence

Apache 2.0

Regions

China

Vendor

Qwen

Provider model key

qwen/qwen3.6-flash-2026-04-16