Qwen3-Next-80B-A3B-Instruct

Qwen3-Next-80B-A3B-Instruct is the first in the Qwen3-Next series, featuring groundbreaking architectural innovations. It uses Hybrid Attention combining Gated DeltaNet and Gated Attention for efficient ultra-long context modeling, High-Sparsity MoE with 512 experts (10 activated + 1 shared) achieving extreme low activation ratio, and Multi-Token Prediction for improved performance and faster inference.