Alibaba has launched a new generation of AI models called Qwen3-Next, designed for high performance with low computational cost. Crystal Liu writes for Alibaba that the new architecture uses several innovations to achieve this efficiency. The first model in the series, Qwen3-Next-80B, is now open source.
According to the company, this 80-billion-parameter model only activates 3 billion parameters during use. This approach, known as a sparse Mixture of Expert (MoE) architecture, significantly reduces computing requirements. Alibaba claims the model surpasses its previous 32-billion-parameter model while using less than 10% of the training cost. For tasks involving large amounts of text, it allegedly delivers more than 10 times the throughput. The model also supports a large context window of 256,000 tokens, which can be extended to one million.
Alongside the new architecture, Alibaba released Qwen3-ASR-Flash, a new model for automatic speech recognition. It supports 11 languages and various dialects and is designed to perform well in noisy environments, even transcribing song lyrics with background music.
The company also previewed its largest model, Qwen3-Max, which has over one trillion parameters. This model is said to follow complex instructions with greater reliability and to have a significantly reduced rate of generating false information. It supports over 100 languages and is optimized for advanced applications.