Alibaba Launches Qwen4-Max with Architecture Upgrades

Alibaba Launches Qwen4-Max with Architecture Upgrades
Alibaba introduced Qwen4-Max, its most capable large language model to date, featuring a new Qwen4 architecture and upgraded attention mechanism. The model has a 2.4 trillion parameter backbone, activating 71 billion at a time, plus 328 billion N-gram parameters that cache repeated calculations. A four-channel residual system with gradient gates improves training efficiency. Attention optimizations include Gated DeltaNet and QSA to reduce hardware demands. Compared to its predecessor, Qwen4-Max scored higher on internal benchmarks and used up to 35% fewer tokens, rising to 49% on coding tasks. It is available now via Alibaba's cloud platform.
Read the original article →