Alibaba launched this Monday (16) the Qwen4-Max, its new multimodal artificial intelligence model, with 2.4 trillion parameters and a context window of 1 million tokens. The technology is already available to developers and companies through API and platforms in the Chinese company's ecosystem.

The model combines text generation, image and video understanding, and advanced reasoning capabilities. Its access was made available on QwenCloud, through API and the Token Plan, in addition to the QwenWork and QwenStudio platforms. Alibaba Cloud Model Studio documentation also confirms Qwen4-Max's availability in different regions.

The architecture uses the Mixture-of-Experts (MoE) system, which activates only a portion of the parameters during processing. Of the 2.4 trillion total parameters, 71 billion are activated per token, an amount 25% lower than that of Qwen3.8-Max, which has the same total scale.

According to Alibaba, this change helps reduce computational consumption without compromising performance. In Work360, a proprietary benchmark that evaluates professional tasks across 721 specialized areas, the new model used 35% fewer tokens and had an average cost per task approximately 60% lower than that of its predecessor, in high-intensity reasoning mode.

The results reflect the conditions of the tests disclosed by the company and do not represent guaranteed savings in real applications. Alibaba also reported that Qwen4-Max reached 6th place in Code Arena, a ranking focused on evaluating models in programming tasks.

Qwen4-Max was not launched with open weights. The company announced that it intends to make smaller versions of the Qwen4 family available in that format, but did not set a specific date.

Alibaba strengthens safeguards, but safety tests still require independent validation

In addition to the architecture changes, Alibaba announced measures to limit misuse of the model in cybersecurity activities. During training, the company says it filtered content related to offensive attacks, preserving knowledge intended for identifying and correcting vulnerabilities.

The system also received specific datasets and post-training adjustments to strengthen protections against chemical, biological, radiological and nuclear risks.

In results disclosed to Reuters, Alibaba said that the model's score on a secure code generation test rose from 8%, recorded by its predecessor, to 14.7%. The company also stated that internal evaluations found no evidence that Qwen4-Max could significantly expand users' ability to carry out dangerous tasks. These conclusions were not independently verified by Reuters.

The announcement comes amid questions about the transparency of safety tests for Chinese models. A survey by consultancy SemiAnalysis, released in October, examined 857 releases from nine Chinese developers, between 2021 and September 2026, and identified public safety evaluation results associated with only 31 of them, equivalent to 3.6%.

Only nine releases, or 1.1% of the total, presented these results by the availability date. The research measured public disclosure of evaluations, not the conduct of internal tests, which may have occurred without publication of results.

The new safeguards announced by Alibaba represent measures declared by the developer itself. Their effectiveness against real attempts at misuse still depends on evidence and external evaluations.

More from Radar