Alibaba launches Qwen4-Max on its cloud platform


Alibaba says the new model scored 52.69 against its predecessor’s 47.74 on its Work360 professional-task test, using 24% to 35% fewer tokens depending on the reasoning setting.
Alibaba has launched Qwen4-Max, the first model in its new Qwen4 family, and made it available through QwenCloud. The release follows Qwen3.8-Max, for which Alibaba published downloadable model weights, the numerical values that determine how a model processes information.
Qwen4-Max accepts text and images and has a context window of 1 million tokens: the amount of information it can process in a single interaction. Tokens are pieces of text used by the model.
Alibaba says Qwen4-Max uses a mixture-of-experts design, which selects parts of a model for each piece of input rather than using the whole model each time. It activates 71 billion of its 2.4 trillion backbone parameters per token. Parameters are numerical values learned during training.
The model also changes its attention mechanism, which helps it process a user’s input. As SiliconANGLE explains, one component is intended to reduce the hardware needed for large prompts, while another skips information not needed for a response. Both are intended to lower processing costs.
The Work360 comparison comes from Alibaba’s own test of professional tasks. Alibaba says Qwen4-Max scored 52.69 against Qwen3.8-Max’s 47.74, using 24% to 35% fewer tokens depending on the reasoning setting. The company also found that Qwen4-Max used 49% fewer tokens than its predecessor on Terminal-Bench 4.0, a suite of coding tasks performed in a terminal environment.
In a separate independent evaluation, Artificial Analysis gave Qwen4-Max 43 points on its Intelligence Index and calculated an average cost of $3.27 per task. Those results are distinct from Alibaba’s Work360 test.
Developers can use Qwen4-Max through hosted services or third-party providers. Alibaba plans to release open weights for smaller Qwen4 models, but has not specified which models or when.
In its November 16 initial findings, Microsoft said increased gateway traffic and reduced server capacity triggered a defect from July, causing requests to pile up until servers stopped responding.
2 min

Barracuda says an unretired vendor access key enabled the intrusion, and identifies ARTEX as the superintelligence (SI) testing framework used in it.
2 min

Z.ai says GLM-5.3-flashX-Legal can help review contracts, analyze evidence and research compliance. The company announced its availability on 13 November 2026, and invites prospective users to download AutoClaw.
1 min