Qwen Releases 80B Mixture-of-Experts Model

The Qwen team at Alibaba has released Qwen3-Next-80B-A3B-Instruct, a new large language model that employs a Mixture-of-Experts (MoE) architecture. This release marks the introduction of the Qwen3-Next series, signaling a focus on more computationally efficient designs for powerful models.

The key feature of this model is its MoE structure. While it contains a total of 80 billion parameters, only 3 billion are activated for processing any given token. This design aims to provide the knowledge and nuance of a very large model while keeping inference costs significantly lower, making it more accessible for a wider range of applications and hardware setups.

Technical Specifications

Beyond its architecture, Qwen3-Next is an instruction-tuned model designed for chat and task completion. It supports a context length of up to 65,536 tokens, making it suitable for tasks requiring long-form context and analysis. The model is built on a standard Transformer foundation with SwiGLU activations and Group Query Attention for efficiency.

Released under the permissive Apache 2.0 license, the Qwen3-Next-80B-A3B-Instruct model is available for both research and commercial use. This continues the trend of major AI labs contributing powerful, open models that allow developers to build without restrictive licensing, fostering broader innovation in the ecosystem.