Qwen3-Next 80B-A3B Instruct

Qwen3-Next-80B-A3B-Instruct is the first release built on the Qwen3-Next architecture developed by Alibaba Cloud's Qwen team. It introduces a hybrid attention system that interleaves Gated DeltaNet linear attention in roughly three quarters of its layers with gated full attention in the remainder, an arrangement designed to keep long-sequence inference efficient while preserving strong retrieval accuracy where pure linear attention tends to weaken. On top of that attention design, the model applies an ultra-sparse Mixture-of-Experts configuration with 512 experts, activating only a small routed subset per token so that the active compute footprint is far lower than the headline parameter count suggests.

Positioned as the instruction-tuned sibling of the Qwen3-Next line, the model is aimed at practical assistant and agent workloads rather than chain-of-thought reasoning. In head-to-head benchmark aggregation it sits slightly below the matching Qwen3-Next-80B-A3B-Thinking variant on composite score, while offering more predictable pricing for production deployments. The open-weight availability makes it well-matched to teams that want to self-host long-context assistants, build retrieval-heavy pipelines, or run tool-using agents on infrastructure where the sparse expert activation and hybrid attention scheme can be fully exploited.