Set Trending Topics as a preferred source on Google.

Bigger does not automatically mean better: Alibaba has unveiled Qwen4-Max, its largest A.I. model yet, with 2.4 trillion parameters and a context window of one million tokens. The model is meant to work on projects around the clock as an “A.I. teammate,” from 3D models and games to legal reviews, the Qwen team wrote in its launch blog. In independent rankings, however, Qwen4-Max lands well behind the top models from the United States and also behind its Chinese rival Moonshot.

What Qwen4-Max Can Do

Qwen4-Max is a so-called mixture-of-experts model: Of its 2.4 trillion parameters, only 71 billion are active per token, plus 328 billion parameters for a new N-gram embedding. Besides text, the model also understands images, and it is a reasoning model. According to Alibaba, it uses significantly fewer tokens for the same tasks than its predecessor, Qwen3.8-Max. On the company’s own Work360 benchmark, the cost per task fell from $4.43 (about 3.98 euros) to $1.67 (about 1.50 euros).

Alibaba presents a whole series of real-world examples:

Industry: At two automotive suppliers, the model reportedly cut the programming effort for a single workpiece from about 475 minutes to 24.
Law: For pre-I.P.O. legal due diligence, Qwen4-Max sorted 198 working papers and delivered an 86-page report.
Research: In a 48-hour autonomous run, the model designed a new coolant for data centers, which an automated lab then synthesized.

Qwen4-Max is available through QwenCloud via API and in Alibaba’s QwenWork, Qoder and QwenStudio apps.

Open Weights Only for the Smaller Models

Anyone hoping to run Qwen4-Max on their own servers is out of luck: Artificial Analysis lists the model as proprietary, and its weights are not available for download. Alibaba is thus sticking with its practice of offering its Max models only through its own cloud, as it has since Qwen2.5-Max. The smaller models in the Qwen4 family, by contrast, are to be released as open source, built on the same architecture and trained with the same techniques, the Qwen team said.

With its open models, Alibaba has built a large developer community and recently even pushed Tencent to follow suit. Jensen Huang recently underlined how important Chinese open-weight models have become: According to the former Nvidia chief, 80 percent of U.S. start-ups use Chinese open models. The strongest open-weight model outside China is currently Inkling 2 from Mira Murati’s Thinking Machines.

Ninth Among the Leaders

Independent tests paint a more sober picture. On the Intelligence Index from Artificial Analysis, Qwen4-Max scores 43 points. Among the 10 leading models the service compares directly, that is only enough for ninth place:

Rank
Model
Developer
Intelligence Index

1
Claude Fable 5.2
Anthropic
65

2
Gemini 4 Carbon
Google
60

3
Claude Opus 5.5
Anthropic
59

4
GPT-6 Astra
OpenAI
57

5
GPT-6.1 Sol
OpenAI
55

6
Muse Wonder Full
Meta
50

7
Kimi K3.1 Pro
Moonshot AI
46

8
Muse Wonder
Meta
44

9
Qwen4-Max
Alibaba
43

10
Grok 4.7
xAI
35

In the intelligence ranking for its model class, Artificial Analysis places Qwen4-Max 38th out of 81, just above the median of 41 points. That leaves Moonshot, with Kimi K3.1 Pro, as the best-performing Chinese developer.

Ahead in Law and Agents, Behind in Coding

Even in Alibaba’s own comparisons, Qwen4-Max does not lead everywhere. The model is strong at legal agent work and long tasks, while Anthropic, OpenAI and Moonshot are ahead in agentic coding. A selection of the figures Alibaba published:

Benchmark
Qwen4-Max
Claude Opus 5.5
GPT-6.1 Sol
GPT-6 Astra
Kimi K3.1

HLAB-Vals 1.1 (legal)
21.3
9.2
6.3
7.1
11.7

DeepSWE v1.1 (coding)
75.7
74.2
75.2
74.1
75.5

Terminal-Bench 4.0
48.5
59.6
56.1
59.1
67.2

ProgramBench
31.5
65.0
52.5
55.0
25.0

Humanity’s Last Exam
51.7
59.8
54.0
53.6
51.1

APEX-Agents v1.1
72.1
73.5
60.0
64.7
65.5

Cheap but Chatty

On price, Alibaba plays to its strengths. According to Artificial Analysis, Qwen4-Max costs $1.50 (about 1.35 euros) per million input tokens and $4.50 (about 4.04 euros) per million output tokens, with a 95 percent discount on cached inputs. A task from the Intelligence Index costs an average of $3.27, compared with $10.04 for Claude Opus 5.5; only GPT-6.1 Sol is cheaper, at $1.27. In return, Qwen4-Max talks a lot: It generated 127 million output tokens for the test, far more than the median of 78.6 million.

Alibaba does not disclose what hardware it used to train the model. The question is sensitive since a forensic analysis found traces of banned Nvidia chips in the newest models from DeepSeek, Moonshot, MiniMax and Alibaba’s affiliate Ant Group.


Rank My Startup: Erobere die Liga der Top Founder!