Qwen3-Next-80B-A3B-Instruct

Qwen3-Next-80B-A3B-Instruct is the first model in the Qwen3-Next series with breakthrough architectural innovations. Uses hybrid attention (Gated DeltaNet + Gated Attention) for efficient ultra-long context modeling, MoE with high sparsity (512 experts, 10 active + 1 shared), and multi-token prediction. 80 billion parameters (3 billion active), trained on 15T tokens. Outperforms Qwen3-32B-Base at 10% of the training cost. Context support up to 256K (expandable to 1M with YaRN). Apache 2.0 license.