Qwen3 Next 80B A3B Instruct (Vertex AI (OpenAI-compatible))

Qwen3-Next-80B-A3B-Instruct is the first release in the Qwen team's Qwen3-Next series, a new foundation-model line aimed at making very large models more efficient to train and serve. The creators frame it as a step toward more agentic AI by improving scaling efficiency through architectural changes rather than simply enlarging parameter counts. As the instruction-tuned variant in that series, it is positioned for practical deployment in chat, reasoning, and long-context workflows where throughput and cost matter as much as raw capability.

The model's signature architectural choices center on a Hybrid Attention design that replaces standard attention with a combination of Gated DeltaNet and Gated Attention, enabling efficient context modeling at ultra-long sequence lengths. It pairs this with a high-sparsity Mixture-of-Experts layout that keeps an extreme low activation ratio, reducing compute per token while retaining large total model capacity, and adds stability optimizations such as zero-centered, weight-decayed layer normalization for more robust pre-training and post-training. Together these features target the same practical sweet spot as much larger dense predecessors, making Qwen3-Next-80B-A3B-Instruct well suited to long-document analysis, retrieval-heavy assistants, and other workflows where very large context windows and steady inference speed are the priority.