Alibaba launches Qwen4-Max AI model to rival US competitors
The new 2.4-trillion-parameter model aims to outperform GPT and Claude in coding and scientific research while significantly cutting operational costs.

The short version
- Qwen4-Max features 2.4 trillion parameters using a Mixture-of-Experts design to activate only necessary portions of the model.
- The model achieved a coding score of 75.7, surpassing Anthropic Opus 5.5 and OpenAI GPT-6.1-Sol in Alibaba's internal benchmarks.
- Alibaba reported a 62% drop in average task costs on its Work360 benchmark, with prices falling from $4.43 to $1.67.
- The model includes specific safeguards against cyberattacks and CBRN-related risks, such as chemical or nuclear threats.
- While Qwen4-Max is closed-source, Alibaba plans to release smaller versions of the Qwen4 family as open weights.
Why it matters
This launch signals a shift toward high-efficiency, heavy-duty AI that can perform specialized scientific and professional work. For users, it means more affordable access to high-performing models that can handle complex tasks like debugging code or discovering new molecules.
A new flagship for heavy tasks
Alibaba Group has officially launched Qwen4-Max, its newest flagship artificial intelligence model. The company aims to strengthen its global position against major US rivals like OpenAI and Anthropic by providing a tool for complex professional work. This new model is specifically designed to handle demanding tasks such as writing and debugging code, analyzing financial documents, and building games.
The model is built on the latest Qwen4 architecture and features a massive scale of 2.4 trillion parameters. To manage this size efficiently, it uses a Mixture-of-Experts (MoE) design. This architecture allows the system to engage only the specific portions of the model required for a particular task, rather than activating the entire network at once.
Beyond text, Qwen4-Max includes native vision capabilities to process visual information alongside written data. It also supports a large context window of up to one million tokens. This allows the model to process and remember vast amounts of information during a single interaction, making it suitable for long-form research and deep data analysis.
Beating the competition in coding
Alibaba claims that Qwen4-Max can match or even surpass the performance of top US AI models in specific categories. According to benchmarks reported by the company, the model achieved a score of 75.7 in coding evaluations. This figure outperformed Anthropic's Opus 5.5, which scored 74.2, and OpenAI's GPT-6.1-Sol, which scored 75.2.
While these results are based on Alibaba's own reported benchmarks and have not been independently verified, they highlight the model's focus on technical proficiency. The company specifically designed the model to assist with software vulnerability detection and remediation. It is intended to provide high-level support for developers while preventing responses that could assist in user-initiated cyberattacks.
To ensure safety in these areas, Alibaba filtered parts of its training data to remove offensive cyberattack capabilities. The company also introduced specific datasets and post-training protections to mitigate risks involving chemical, biological, radiological, and nuclear (CBRN) applications. These safeguards are part of a broader effort to manage the threats posed by increasingly sophisticated AI systems.
Efficiency and cost reductions
One of the most significant updates with Qwen4-Max is its improved efficiency compared to its predecessor, Qwen3.8-Max. Alibaba reports that the new model is 35% more efficient in terms of tokens on its Work360 benchmark. It also demonstrated a 49% increase in efficiency on the Terminal-Bench 4.0 benchmark.
These efficiency gains translate directly into lower costs for users who rely on the model for high-volume tasks. Alibaba reported a 62% drop in the average cost of tasks on the Work360 benchmark. Specifically, the average cost for these tasks decreased from $4.43 to $1.67, making the model more accessible for large-scale enterprise applications.
The model is available through several of Alibaba's cloud platforms, including QwenCloud, QwenWork, Qoder, and QwenStudio. For those looking to build applications, the API pricing for cached inputs starts at $0.075 per million tokens. This pricing structure aims to provide a balance between high-end performance and economic viability for developers.
Scientific research and discovery
Alibaba is positioning Qwen4-Max as a tool for advanced scientific discovery rather than just a general-purpose chatbot. The company shared that the model ran on a continuous 48-hour research session to discover a new molecule. This molecule was experimentally verified as a coolant for the next generation of semiconductor chips.
In addition to hardware-related research, the company announced an application for the model to perform autonomous research on Alzheimer's disease. These use cases demonstrate the model's ability to handle complex, multi-step reasoning across different scientific disciplines. It serves as a specialized assistant for researchers who need to process large amounts of data to find specific results.
To help users move from ideas to prototypes, Alibaba also launched a feature called Live Delivery. This system creates an initial working prototype of a project and then refines it through subsequent iterations based on user feedback. To encourage broader adoption, the company is making the Qwen-Live-Delivery-Plugin associated with this system open source.
Open weights and closed models
The launch of Qwen4-Max comes with a specific strategy regarding the accessibility of Alibaba's AI technology. While Qwen4-Max itself remains a closed-source model, Alibaba has announced that smaller models within the Qwen4 family will be released as open weights. This allows the company to maintain its flagship model's proprietary status while still contributing to the open-source AI community.
By offering smaller models as open weights, Alibaba enables developers to build and customize their own tools without the restrictions of a closed API. This move is intended to keep the company's open-source AI platform growing while protecting the intellectual property of its most powerful flagship. It creates a tiered ecosystem where different sizes of models serve different needs.
Moving forward, the availability of these models will likely influence how developers choose their infrastructure. Users who need the highest level of performance for complex tasks can opt for the closed Qwen4-Max. Meanwhile, those looking for flexibility and customization can utilize the smaller, open-weight versions of the Qwen4 family to build specific applications.
What to watch
- Independent verification of Alibaba's coding benchmarks against GPT and Claude.
- The release dates for the smaller Qwen4 open-weight models.
- Further results from the autonomous Alzheimer's research project.
Try it
Check out the QwenCloud or QwenStudio platforms to see the current API pricing for Qwen4-Max tokens.
Sources and links
- ReportingMint on MSN: the article we sourced
First reported by Mint on MSN. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.