Qwen3 Next 80B A3B Instruct
Alibaba · 2025-09-10 · 81.3B parameters
Qwen3-Next-80B-A3B-Instruct is Alibaba's instruction-tuned, non-thinking text model built on a highly sparse mixture-of-experts architecture; Qwen describes it in rounded terms as 80B total parameters with 3B activated, while the released checkpoint contains 81,324,862,720 parameters. It combines Gated DeltaNet with gated grouped-query attention and supports 262,144 tokens natively, with documented YaRN extension to 1,010,000 tokens.