AI Agent News/stories/alibaba-releases-qwen4-max-with-cheaper-training-and-a-new-attention-mechanism.md

InfrastructureBreaking

Alibaba releases Qwen4-Max with cheaper training and a new attention mechanism

Alibaba has introduced Qwen4-Max, its most capable large language model, which it says produces better output at lower cost than its predecessor. The model is available through Alibaba's cloud platform, and the company is also open-sourcing a plugin aimed at long-running tasks.

Source: siliconangle.com/2026/11/16/alibaba-launches-qwen4-max-with-training-attention-optimizations

Alibaba says Qwen4-Max is built on a new Qwen4 architecture that makes training more efficient. The backbone has 2.4 trillion parameters, of which 71 billion are active at any given time, and it is supported by 328 billion N-gram parameters that act as a cache for frequently repeated calculations.

The model also changes how gradients move through its layers. Qwen4-Max uses four residual channels instead of one, with gates that adjust gradient flow during training. Its attention mechanism adds Gated DeltaNet, which lowers the hardware needed for large prompts, and QSA, which skips data the model does not need to answer a prompt.

In an internal benchmark covering knowledge work, Alibaba reports that Qwen4-Max scores higher than Qwen3.8-Max while using up to 35% fewer tokens. On the Terminal-Bench 4.0 coding suite the token saving reaches 49%. Alibaba also plans to open-source smaller models built on the same technologies.

Why it matters if you build agents

Developers running long agent tasks on Qwen4-Max may see lower token costs per task, though the figures come from Alibaba's own tests. The Qwen-Live-Delivery-Plugin returns an early draft before a long run finishes and accepts feedback along the way, which lets agents be reviewed and corrected mid-task. Check these claims against your own workloads before switching models.

Get stories like this every weekday

One short email with the day's agent tooling news.

Subscribe