Qwen3-Next: A New Blueprint for Efficient AI

Alibaba's Qwen3-Next models represent a significant shift in building large language models (LLMs) by prioritizing efficiency and architectural innovation over sheer size. With a design that activates only 3 billion out of its 80 billion parameters for any given task, the Qwen3-Next models utilize an ultra-sparse Mixture of Experts (MoE) architecture, which directs tasks to a select few specialized experts, resulting in models that are 10 times cheaper and faster than their predecessors. These models also introduce a hybrid attention mechanism for processing long documents, combining a linear Gated DeltaNet for speed and a Gated Attention mechanism for precision, enabling them to handle long texts efficiently. Released under the Apache License, Version 2.0, these models offer unmatched efficiency and top-tier performance, with integration capabilities via Alibaba Cloud Model Studio and other platforms, setting a new standard for future AI developments by focusing on smarter, rather than larger, architectures.