Alibaba says Qwen4-Max cuts task costs by 60%, boosts cyber safeguards
Alibaba on Monday launched Qwen4-Max, its latest flagship artificial intelligence (AI) model, claiming it reduces average task costs by around 60 per cent while strengthening safeguards against cybersecurity threats and other risks.
The company said the model outperforms its predecessor, Qwen3.8-Max, on complex professional tasks while using 35 per cent fewer tokens in high-effort reasoning mode.
According to the company's press release, Qwen4-Max has 2.4 trillion parameters, native vision capabilities and a context window of one million tokens.
Alibaba described it as the "most capable model in the Qwen family to date".
Model architecture and benchmark performance
Qwen4-Max is built on an upgraded sparse Mixture-of-Experts (MoE) architecture, which allows it to activate only a portion of its total parameters for each token processed, the company said.
Also Read
Although the model has 2.4 trillion total parameters, it activates 71 billion parameters per token, 25% fewer than Qwen3.8-Max at the same overall scale, according to Alibaba.
The company said Qwen4-Max outperformed Qwen3.8-Max on its proprietary Work360 benchmark, which evaluates performance across 721 specialised fields.
In high-effort reasoning mode, the model used 35% fewer tokens and reduced the average cost per task by around 60 per cent.
Qwen4-Max also ranked sixth on Code Arena, a real-world coding leaderboard.
Alibaba reported strong results on other benchmarks covering agent workflows, workplace productivity and visual reasoning, including Agents' Last Exam, Vals Index, JobBench and Chartography.
Applications in professional workflows
The model is designed to handle tasks across 3D modelling, game development, precision manufacturing, finance, law and healthcare.
Alibaba said Qwen4-Max also supports cloud resource scheduling. In benchmark testing, its scheduling capabilities reduced data-annotation model call costs by an average of 9% and increased offline CPU rescheduling speeds by more than 14 times.
In automotive manufacturing, the model can assist with production-line planning, part design and machining verification. During a company demonstration, it reduced manual workpiece programming time by 95 per cent, from around 475 minutes to 24 minutes, by identifying excessive cutting and insufficient material removal before machining began.
Scientific research applications
Alibaba also reported using Qwen4-Max for an experiment involving Alzheimer's research. The model handled tasks covering target research, molecular design and computational evaluation.
During the workflow, it autonomously executed 454 tool calls and 3,239 metric evaluations to design and optimise 24 candidate molecules, according to the company.
Cybersecurity and safety measures
Alibaba said it incorporated security checks into software engineering training alongside functionality tests to address vulnerabilities such as code injection, memory-safety flaws and broken access controls.
The company also filtered training data according to task intent. It retained defensive cybersecurity material used for vulnerability detection and remediation while removing content related to offensive cyberattacks.
In addition, Alibaba introduced dedicated safety datasets and post-training improvements to strengthen protections against chemical, biological, radiological and nuclear (CBRN) risks.
According to Reuters, Alibaba said its internal testing had found no evidence of significant high-risk capabilities in the model. The company also said it had strengthened safeguards against potential misuse in areas such as cybercrime and hazardous materials.
Availability and pricing
Qwen4-Max is available through Alibaba's QwenCloud platform via its Token Plan and application programming interface (API). Users can also access it through QwenWork and QwenStudio.
Alibaba has priced input tokens with an implicit cache hit at $0.075 per million tokens. The company said smaller, open-weight models from the Qwen4 series would be released to developers worldwide.
What's Your Reaction?
Like
0
Dislike
0
Love
0
Funny
0
Wow
0
Sad
0
Angry
0
Comments (0)