Alibaba's Qwen 4 Max Beats Chinese Rivals on AI Agents, Stays Locked Down
Alibaba's Qwen 4 Max is out: 2.4 trillion parameters, a new Qwen 4 architecture, and the top Chinese model on Arena's agent leaderboard. Its coding and design scores land near the best American models. One catch — it's closed-weight, so you can use it, not download it.
- 2 sources report it: decrypt.co, kucoin.com.
- The source published it on Nov 16, 2026 at 10:46 PM.
Nothing open. Every check passed.
Alibaba dropped Qwen 4 Max on Monday, the first flagship built on its new Qwen 4 architecture and the company's most powerful model yet. It carries 2.4 trillion parameters and handles text, images and video, with space for up to one million tokens of context — enough for hefty codebases or stacks of documents. It is also the highest-ranked Chinese model on Arena's agent leaderboard, ahead of Tencent, Moonshot AI and DeepSeek on that test.
The model's trick is a mixture-of-experts design that keeps most parameters idle for any given word. Qwen 4 Max activates 71 billion at a time, down from 95 billion for Qwen 3.8 Max, letting a huge model spend less compute per token. On Agent Arena it placed 14th among individual models and made Alibaba the fifth-ranked lab overall, posting a preliminary 7.46% score across 14,727 sessions — the best of any Chinese entry. The board judges real agent sessions on task completion, user satisfaction and how well a model takes corrections, scoring each by net improvement over an average competitor.
Visual and coding work may be the bigger draw for regular users. Qwen 4 Max debuted sixth globally on Code Arena's web-development ranking with 1,750 points, 74 more than its predecessor, and took eighth in the branding and reference-based design categories. Alibaba's own demos go further, showing it build game environments and convert factory layouts into 3D models — company-selected examples, not independent tests. The visual capability means understanding images and video and producing code or other text, not releasing a new image generator.
Through QwenCloud, the model costs $1.50 per million input tokens and $4.50 per million generated tokens, a 25% reduction from the previous flagship's $2 and $6 rates. Reused input served from an automatic cache runs $0.075 per million tokens, a cushion for repetitive workflows. Alibaba says its internal testing across hundreds of professional specialties found the model burned 35% fewer tokens and cut average task costs by roughly 60% compared with Qwen 3.8 Max.
Here's the snag for the community: Qwen 4 Max is closed-weight. Developers can reach it through an API or Alibaba's apps, but they can't download its internal parameters, modify it or run it privately on their own hardware. That's a pivot for a family that has become the backbone of the open-model world — an August report from Hugging Face counted 151,448 Qwen-based derivative models, about 4.7 times the number built on Meta's Llama family. Alibaba opened the weights of Qwen 3.8 Max earlier this year and had previewed Qwen 4 architecture elements with smaller models months before Monday's launch.
The launch lands in a crowded Chinese race. Moonshot's Kimi K3 pushed enormous open-weight models toward Western performance, and DeepSeek keeps swinging 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 charge a lot more. Alibaba says smaller Qwen 4 models will arrive with open weights, so developers can build on the generation rather than just rent its biggest model. It hasn't announced their sizes or release dates.
TopicsAlibaba · Qwen 4 Max · Qwen 3.8 Max · Arena · Moonshot AI · DeepSeek · Tencent · Hugging Face
Written by AI from verified reporting.

