Qwen3 Next 80B A3B Thinking

Qwen3 Next 80B A3B Thinking is a specialized reasoning model built on the innovative Qwen3-Next architecture, which prioritizes extreme efficiency in both training and inference. The model employs a hybrid attention mechanism that combines Gated DeltaNet and Gated Attention to manage long-context modeling effectively. By utilizing a high-sparsity Mixture-of-Experts structure, it achieves a low activation ratio, allowing the 80-billion-parameter model to activate only 3 billion parameters during inference. This design, complemented by multi-token prediction to accelerate generation, makes it particularly well-suited for demanding tasks such as mathematical proofs, code synthesis, and complex agentic planning.

The model benefits from advanced stability optimizations, including zero-centered and weight-decayed layernorm, which ensure robust performance during pre-training and post-training phases. By leveraging GSPO, the development team successfully addressed the stability challenges inherent in combining hybrid attention with high-sparsity MoE during reinforcement learning. As a reasoning-first chat model, it is tuned to output structured thinking traces, providing a reliable framework for retrieval-heavy workflows and multi-step logic. Its ability to maintain stability across long chains of thought while reducing off-task behavior positions it as a powerful tool for developers building sophisticated agent frameworks and standardized benchmarking applications.