Alibaba has announced the launch of Qwen4-Max, a new frontier model which the firm claims can compete with leading Western offerings as well as bringing cyber safety and token efficiency improvements.
The Chinese tech giant said its new 2.4 trillion model is a major improvement over its predecessor Qwen3.8-Max across coding, science and industry-specific tasks.
Amid warnings that AI models are accelerating cyber risks, Alibaba also said the model produces safer code and has more rigorous safety controls to prevent potential misuse.
In its own benchmark Last Secure Code Bench, which tests models on their capacity to generate secure code, Qwen4-Max scored 31.8 per cent, a 9.8 per cent rise compared to its predecessor.
Alibaba added that internal safety testing shows no large-scale cybersecurity capabilities for the model and used post-training to increase the model’s refusal of unsafe requests.
It also claimed that Qwen4-Max’s token usage is more efficient, with an average usage drop of 24 per cent at its highest effort setting.
Paired with lower API prices at launch, $1.413 per million input tokens and $4.24 per million output tokens, Alibaba said the total cost per task is 62 per cent lower than Qwen3.8-Max.
In its own tests, Alibaba found Qwen4-Max outperformed OpenAI’s GPT-6 Astra in benchmarks such as JobBench, Automationbench-AA and HLAB-Vals-1.1, which test AI models across real-world professional tasks, applications and legal work.
Alibaba added that the model represents a leap in multimodal and image analysis capabilities. It gave the example of Chinese automotive manufacturers already using Qwen4-Max for part design, production-line planning and live identification of issues such as overcuts using multimodal input.
Additional testing found Qwen4-Max performed almost ten points better than its predecessor in the agentic coding benchmark Terminal-Bench 4.0, scoring 48.5 per cent, while still behind models Claude Opus 5.5, GPT-6 Astra and domestic rival Moonshot AI’s Kimi K3.1 Pro.
The third-party AI benchmarking platform Artificial Analysis said Qwen4-Max is “above average in intelligence” and scored 43 on its intelligence index, which tests AI models in nine benchmarks covering coding, knowledge work and science.
This puts the model ahead of the 42 scored by Anthropic’s Claude Sonnet 5.5 and just behind Meta’s Muse Wonder at 44, while trailing Kimi K3.1’s score of 46.
In testing, Alibaba added, Qwen4-Max has reduced the time it takes to program a manufacturing line to produce a specific component from 475 minutes to 24 minutes.
Alibaba has moved to increase its share of the Chinese AI market in recent months. Last week, it announced plans to list its semiconductor subsidiary T-Head, which produces the chips used to train and inference Qwen models, on the Hong Kong Stock Exchange.








Recent Stories