Alibaba's Qwen 4 Max Leads China's AI Agent Race, but Its Weights Remain Closed
Key Takeaways
- •Qwen 4 Max carries 2.4 trillion parameters, supports a one-million-token context window, and natively handles text, images, and video.
- •The model is the highest-ranked Chinese entry on Arena's agent leaderboard, placing 14th among individual models with a preliminary 7.46% score across 14,727 sessions.
- •Qwen 4 Max costs $1.50 per million input tokens and $4.50 per million generated tokens, a 25% reduction from its predecessor, and Alibaba's internal testing showed roughly 60% lower average task costs.
- •The flagship is closed-weight, meaning developers can only use it through Alibaba's API or apps and cannot download or run it on their own hardware.
- •Hugging Face counted 151,448 Qwen-based derivative models, about 4.7 times the number built on Meta's Llama family, highlighting Qwen's role as a foundation of the open-model ecosystem.

Alibaba released Qwen 4 Max on Monday, its most powerful AI model to date and the first flagship built on the company's new Qwen 4 architecture. According to Alibaba's official announcement and its Qwen blog post, the model carries 2.4 trillion parameters, accepts up to one million tokens of context, and natively processes text, images, and video.
It is also the highest-ranked Chinese model on Arena's agent leaderboard, putting Alibaba ahead of Tencent, Moonshot AI, and DeepSeek in that particular test.
Parameters are, in effect, all the different dials a model can adjust when processing information, while tokens are the smallest units of information a model handles. Alibaba uses a mixture-of-experts design, which keeps most of those parameters idle for any given word the model processes. Qwen 4 Max activates 71 billion parameters at a time, down from 95 billion for its predecessor, Qwen 3.8 Max, allowing a very large model to use less computing power per token. The one-million-token context window provides enough room for lengthy codebases or stacks of documents.
Top Chinese Entry on Agent Arena
On Agent Arena, Qwen 4 Max placed 14th among individual models and made Alibaba the fifth-ranked laboratory overall. The leaderboard evaluates real agent sessions using task completion, user satisfaction, and how well models respond to corrections, then scores each model by its net percentage improvement over an average competitor. Agents — models that carry out multi-step tasks on a user's behalf, calling on tools as they go rather than answering a single prompt — are increasingly the format in which model output reaches working software, which is why these standings draw attention from developers choosing what to wire into their products. Qwen 4 Max posted a preliminary 7.46% score across 14,727 sessions, the highest among Chinese entries; because the figure is preliminary, its position can shift as more sessions are recorded.
Coding Skill and an Eye for Design
The model's visual work may be the more interesting part for everyday users. The Code Arena web-development ranking compares websites built by AI models using head-to-head user votes and a points-based rating system. Qwen 4 Max debuted sixth globally at 1,750 points, up 74 from its predecessor, and placed eighth in the categories for branding and reference-based design.
That puts the model in the same neighborhood as the best American systems at turning a request, a screenshot, or a rough idea into a working interface that actually looks good. Alibaba's demos go further, showing the model building game environments and translating factory layouts into 3D models, although those are company-selected examples rather than independent tests.
The visual capability applies to understanding images and video and producing code or other text-based outputs. It is not the same as releasing a new image generator.
Cheaper Intelligence, With a Catch
Qwen 4 Max costs $1.50 per million input tokens and $4.50 per million generated tokens through QwenCloud, a 25% reduction from the previous flagship's $2 and $6 rates. Reused input served from an automatic cache costs $0.075 per million tokens, which can make repetitive workflows substantially cheaper — a discount that can compound in agent-style applications, which resend the same long context at each step of a multi-step task.
Alibaba says its internal testing across hundreds of professional specialties found that the model used 35% fewer tokens and cut average task costs by roughly 60% compared with Qwen 3.8 Max. In other words, executing tasks on the model should be cheaper even against competitors with similar list prices.
The bigger limitation for the developer community is that Qwen 4 Max is closed-weight. Developers can use it through an API or Alibaba's apps, but they cannot download its internal parameters, modify them, or run the model privately on their own hardware. In practice, that ties any use of the flagship to Alibaba's hosted platform, including its pricing and availability.
A Pillar of the Open-Model Ecosystem
Qwen has become a foundation for the open-model ecosystem. An August report from Hugging Face counted 151,448 Qwen-based derivative models on its platform, about 4.7 times the number built specifically on Meta's Llama family. Those include community fine-tunes and optimized versions that power local assistants and specialized tools.
Alibaba opened the weights of Qwen 3.8 Max earlier this year, despite having kept some previous Max models proprietary, and previewed elements of the Qwen 4 architecture with smaller models months before Monday's release.
The launch lands in a crowded Chinese race. Moonshot's Kimi K3 pushed enormous open-weight models toward Western performance, while DeepSeek has been competing hard on API prices. Qwen 4 Max leads some Chinese rankings, but American labs still hold the top spots in Arena's overall agent and web-development tables, and those models are considerably more expensive.
Alibaba says smaller Qwen 4 models will be released with open weights, giving developers a chance to build on this generation rather than simply rent its biggest model. The company has not announced their sizes or release dates.