Alibaba launches Qwen4-Max, scores 43 on Artificial Analysis

AI model illustration: a layered core glowing red with small figures below

Alibaba's Qwen team has released Qwen4-Max, the largest Qwen model to date. It has 2.4 trillion total parameters and uses a mixture-of-experts architecture, with roughly 71 billion parameters active per token, plus a separate 328-billion-parameter N-gram embedding module. The model accepts text and image input, offers a reasoning mode, and, according to the official description, has a context window in the million-token range.

A preview version is first going live on Alibaba's Token Plan, with a limited-time discount of up to 90% off. Qoder, the Qwen App and QwenStudio have already switched to the new model, and the API is offered through Alibaba Cloud. The weights will not be available for download, continuing the practice of keeping the Max series on Alibaba's own cloud, which dates back to Qwen2.5-Max. The team says smaller models in the Qwen4 family will be open-sourced, built on the same architecture and training methods, but it has not given sizes, a license or a release date.

The numbers Alibaba published

In its launch blog post, the Qwen team positions the model as an "AI teammate" that can work continuously, and cites several real-world examples. Two automotive parts suppliers used it to write machining programs, cutting programming time per workpiece from about 475 minutes to 24 minutes. In a pre-IPO legal due diligence review, the model organized 198 working documents and produced an 86-page report. In another case, during a 48-hour autonomous run, the model designed a data-center coolant that an automated lab then synthesized. All of these cases are self-reported by the company; customer names and acceptance criteria have not been disclosed.

On cost, Alibaba says token consumption on comparable tasks is clearly lower than with Qwen3.8-Max. On its own Work360 benchmark, the cost per task fell from $4.43 to $1.67.

Alibaba also published a table of individual results that shows both strengths and weak spots. On the legal-agent benchmark HLAB-Vals 1.1, the model scores 21.3 against 9.2 for Claude Opus 5.5. On DeepSWE v1.1 coding it scores 75.7, bunched between 74 and 76 with Opus 5.5, GPT-6.1 Sol and Kimi K3.1. On Terminal-Bench 4.0 it manages only 48.5, compared with 67.2 for Kimi K3.1, and on ProgramBench it scores 31.5 against 65.0 for Opus 5.5.

Where it lands on third-party rankings

The independent evaluator Artificial Analysis gives Qwen4-Max an Intelligence Index of 43, up 9 points from Qwen3.8-Max's 34. It ranks 38th among 81 models in the same price tier, slightly above the median of 41. Among the ten leading models it is directly compared with, Qwen4-Max places ninth, behind Claude Fable 5.2 (65), Gemini 4 Carbon (60), Claude Opus 5.5 (59), GPT-6 Astra (57) and others. The only Chinese model ahead of it is Moonshot AI's Kimi K3.1 Pro (46).

Another leaderboard flatters it more. According to Arena, Qwen4-Max took first place among Chinese models on Agent Arena, and Alibaba climbed to 14th on the overall leaderboard.

Price is its selling point. Based on the official rates recorded by Artificial Analysis, input costs $1.50 per million tokens and output $4.50, with a further 95% off for cache-hit input. Completing one Intelligence Index task costs an average of $3.27, which is 67% below Qwen3.8-Max's $10.02 and slightly below the $3.68 of Kimi K3.1 Pro. Tencent's Hy4 is cheaper still at about $1.03, though it scores 8 points lower.

The catch is verbosity. During the evaluation Qwen4-Max produced 127 million output tokens in total, against a median of 78.6 million for comparable models, roughly 60% more by a rough calculation. The unit price came down, but higher usage claws part of the saving back.

Side by side with its peers

Lining up the leading Chinese models makes the picture clearer. Kimi K3.1 Pro scores 3 points higher on the Intelligence Index, costs about $0.40 more per task, and is clearly stronger at terminal-style tasks. Qwen4-Max wins on legal work, long-horizon tasks and price. Tencent's Hy4 takes the ultra-low-price route. Of the three, only Qwen4-Max keeps its weights closed, and the developer community Qwen built up through its open-source models is precisely where it differs from the other two.

For developers in China, the practical step for now is to run a round of real tasks during the Token Plan discount period. Max is not open source, so private deployment will have to wait for the smaller models; Alibaba has not said when they will be released or how large they will be. Also, many of the scores in Alibaba's blog post use its own testing methodology, and Artificial Analysis has not yet published per-task latency data, so latency cannot currently be verified.

Sources: Qwen team launch blog, Artificial Analysis model page (Intelligence Index, cost per task, output tokens), CocoLoop, Arena leaderboard, Trending Topics; Cailian Press reporting was used to cross-check availability channels and discounts. Parameters and individual results follow the figures published by Alibaba.