China's Alibaba has launched its new artificial intelligence model, Qwen4-Max, in a move aimed at strengthening its presence in the advanced-model market and competing with models developed by companies such as OpenAI, Anthropic and Google. With the model, the company is focusing on cutting running costs and improving safety standards.
Parameters and Running Costs
Alibaba Cloud said in an announcement published on its official website the day before yesterday that the new model has about 2.4 trillion parameters, but uses only 71 billion parameters for each token it processes, a mechanism intended to make use of the model's capabilities while reducing the computing resources needed to run it.
The company explained that the average cost of completing a single task, according to its internal benchmark "Work360", fell from $4.43 in the previous version, "Qwen3.8-Max", to $1.67 in the new version, a decrease of 62 percent. It added that the model needs up to 35 percent fewer tokens to complete the same task.
Using Qwen4-Max through the API costs about $1.50 per million input tokens and $4.50 per million output tokens, according to data published by the Qwen team, while the prices of competing models vary depending on the version and type of use.
Cybersecurity and Safety
Alibaba stressed that cutting costs is not the only goal of developing the model, as it also focused on strengthening its capabilities in cybersecurity and secure code writing, by filtering training data and retaining material that helps detect and fix security vulnerabilities, while excluding data that could enhance offensive capabilities.
The company noted that coding test results showed an improvement in the model's performance, as its success rate in the "Last Secure Code Bench" test rose from 22 percent to 31.8 percent. It also conducted specialized training to strengthen the model's ability to deal safely with chemical, biological, radiological and nuclear risks.
In its assessment of the safety level, the company said it had not detected capabilities that could pose wide-scale risks to cybersecurity, but these results are based on internal tests, and the model has not yet undergone independent evaluations by teams specializing in simulating cyberattacks.
Availability and Future Plans
The model is available through the API and the QwenCloud platform, in addition to the Qwen Work, Qoder and Qwen Studio applications. The company plans to release smaller open-source versions of the Qwen4 family in the coming period.
Source: Syrian Arab News Agency (SANA)

Comments (0)