Alibaba says its new Qwen4-Max language model cuts task costs by about 60%

Alibaba calls Qwen4-Max the most capable model in its Qwen family. The company says it removed offensive attack material from the model's training data and kept defensive security content.

Alibaba released Qwen4-Max on 16 November 2026. The company calls it the most capable model in its Qwen family so far and says its average cost per task is around 60% lower than that of its predecessor, Qwen3.8-Max1.

According to Alibaba, the model has 2.4 trillion parameters in total but activates only 71 billion for each token, using an upgraded sparse Mixture-of-Experts (MoE) architecture1. In a mixture-of-experts model, each input is routed to only some of the network's specialised sub-networks rather than through all of them. Alibaba says this means 25% fewer activated parameters than Qwen3.8-Max, which is the same overall size1. Qwen4-Max can also process images natively and has a context window of one million tokens1.

The cost figure comes from Work360, a benchmark that Alibaba owns, which tests difficult industry tasks across 721 specialised fields1. The company says Qwen4-Max beat its predecessor overall on that test and used 35% fewer tokens in its High effort reasoning mode13. MLQ.ai reports that on Work360 the per-task average went down to $1.67 from $4.43 with the older model, which it describes as a 62% reduction2. It also reports that input is now listed at $1.50 per million tokens, down from $2, and output at $4.50 per million, down from $62. MLQ.ai notes that these figures have not been checked independently2.

Alibaba also says it tightened security controls during training. For cybersecurity, it filtered training data by the intent of each task, removing material on offensive attacks while keeping material that helps find and fix vulnerabilities1. The company says coding training added security checks alongside functionality tests, and that dedicated safety datasets strengthen protection against chemical, biological, radiological and nuclear risks1.

As an example of long-running agent work, Alibaba says Qwen4 ran for 48 hours without stopping. Over four rounds of a closed-loop workflow, it found a coolant molecule for chip cooling that was then validated by experiment1.

The model is available through the API and Token Plan on QwenCloud, and in Alibaba applications including QwenWork and QwenStudio1. CASRAI reports that Alibaba's changelog lists qwen4-max on 16 November, five days after a preview that was available only through Token Plan3. The flagship model was released without downloadable open weights, CASRAI reports3. Alibaba says it plans to release smaller Qwen4 models with open weights to developers soon, and CASRAI notes that the announcement gives no date for them13.

Sources

  1. 1
  2. 2
  3. 3
    Alibaba releases Qwen4-Max with 1M context

    IndependentCASRAIAccessed

ITANU’s newsroom system produced this article with AI and checked it against the sources listed above. How we work

ITANU Magazine

Technology news, checked against its sources