Alibaba Unveils Advanced Qwen3-Next AI Models on NVIDIA Platform

Alibaba has introduced two new open-source AI models, Qwen3-Next 80B-A3B-Thinking and Qwen3-Next 80B-A3B-Instruct, featuring a hybrid Mixture of Experts (MoE) architecture. These models are designed to enhance efficiency and performance, particularly when deployed on NVIDIA's advanced platform. The architecture activates only 3 billion of the total 80 billion parameters per token, combining the power of large-scale models with the efficiency of smaller ones.

Optimized for long context lengths, the models can process over 260,000 tokens in input. They leverage NVIDIA's Blackwell 5th-generation NVLink, which provides 1.8 TB/s of direct GPU-to-GPU bandwidth, significantly reducing latency and improving token throughput during complex tasks. The models incorporate 48 layers, with every fourth LAYER featuring innovative architectural enhancements.