Alibaba has officially unveiled Qwen4-Max, the first model in its next-generation Qwen4 family, with a focus on complex professional tasks and autonomous AI agents.
The Chinese tech titan teased the new Qwen4 family last week, taking to X to promise improvements in coding, reasoning and real-world knowledge, while initially making Qwen4-Max available in preview through several of its products.
Now, the company has formally detailed the model’s capabilities, pitching it as an always-available “AI teammate” capable of tackling everything from software development to scientific research.
However, Alibaba appears to be keeping its most powerful model proprietary, despite the open-weight releases that have helped establish Qwen as a major force in the AI realm.
Under the hood of Qwen4-Max
In a blog post published on Monday, Alibaba revealed that Qwen4-Max is built on a new mixture-of-experts (MoE) architecture comprising 2.4 trillion backbone parameters, with 71 billion activated for each token. It also includes another 328 billion parameters dedicated to a technique called “N-gram Embedding,” which helps the model process patterns in text more efficiently.
The model supports images as well as text, and comes with a one-million-token context window, allowing it to process large amounts of information in a single interaction.
Qwen4-Max is available through Alibaba’s QwenCloud platform, with API pricing starting at $1.50 per million input tokens and $4.50 per million output tokens. It can also be accessed through QwenWork, Qoder and QwenStudio, while Vercel has added Qwen4-Max to its AI Gateway, allowing developers to use the model through coding tools such as Claude Code, Codex and Cursor.
Alibaba claims the new architecture delivers substantial improvements over its predecessor, Qwen3.8-Max, particularly when performing complex tasks involving multiple tools and extended periods of autonomous execution.
On Work360, an internal benchmark developed to assess professional tasks across hundreds of specialised fields, Qwen4-Max scored 52.69 compared with 47.74 for Qwen3.8-Max. Alibaba also claims the newer model uses between 24% and 35% fewer tokens on that benchmark, depending on the reasoning effort selected.
The company reports even greater token savings on Terminal-Bench 4.0, a test of AI agents’ ability to carry out tasks in a terminal environment, where token usage fell by nearly half compared with Qwen3.8-Max.
Alibaba says those improvements, combined with lower API prices, reduced the average cost of completing a task in its Work360 evaluation by 62%, from $4.43 to $1.67.
Of course, those are Alibaba’s own results. Independent testing by Artificial Analysis offers another perspective, awarding Qwen4-Max a score of 43 on its Intelligence Index, placing it 38th among 81 models in the comparison. That puts it slightly above the median score of 41 for models in its price category, but lagging leading competitors such as Anthropic’s Claude Fable 5.2 (65), Google’s Gemini 4 Carbon (60) and OpenAI’s GPT-6 Astra (57).
Artificial Analysis also calculated an average cost of $3.27 per task across its evaluations, while noting that Qwen4-Max generated 127 million output tokens, considerably more than the median of 78.6 million among comparable models.

Alibaba pulls its weight(s)
The proprietary nature of Qwen4-Max is particularly notable given Alibaba’s recent history. When the company introduced Qwen3.8-Max in August, it made an unusual commitment: for the first time, it would release the weights underlying one of its flagship Max models, allowing developers to download and run a version on their own infrastructure.
That marked a departure from Alibaba’s previous approach, with Qwen3.6-Max-Preview and Qwen3.7-Max both remaining proprietary. In hindsight, Qwen3.8-Max looks more like an exception than the beginning of a broader shift towards open weights.
The announcement followed Alibaba’s efforts to demonstrate that its models could compete with leading proprietary systems from the likes of Anthropic. As The New Stack reported at the time, the company had already been making increasingly ambitious performance claims for its Qwen family.
Qwen3.8-Max also arrived with a demonstration of autonomous software development, in which an AI coding agent worked on a project over 16 days, producing hundreds of commits and more than 100 pull requests.
However, the open-weight promise initially attracted scepticism when Alibaba announced the model release before making the files available.
“That is not a technical footnote; that is the API business model wearing an open source jacket for the launch photo.”
In reactions gathered by The New Stack, independent AI consultant Jeff Brokaw questioned the significance of advertising downloadable weights that developers could not yet obtain. He also drew parallels with Chinese AI startup Moonshot AI’s Kimi K3, which had similarly been promoted as an open-weight model ahead of the files becoming available.
“That is not a technical footnote; that is the API business model wearing an open source jacket for the launch photo,” Brokaw said at the time.
Alibaba, for its part, did eventually follow through on August 12, publishing the underlying Qwen3.8-2.4T-A95B weights — but there were notable strings attached. The downloadable model came under a custom licence, versus the more permissive Apache 2.0 terms Alibaba had used for various earlier Qwen releases. Businesses operating “model-as-a-service” or AI work assistant products, with combined revenue exceeding $50 million over any consecutive 12-month period, were required to obtain a separate commercial licence, although internal use was exempt under specified conditions.
The licence also required certain commercial products with more than 100 million monthly active users, or $20 million in monthly revenue, to prominently display the model’s name.
For comparison, Xiaomi released its MiMo-V2.6 models under the permissive MIT licence in September, alongside technical documentation and a public record of its reinforcement-learning training costs. The company also committed to publishing thousands of training environments and associated tools, giving outside developers a potentially deeper view of how the models were developed.
Alibaba was willing to wear the “open source jacket,” while keeping some of its most valuable capabilities behind a commercial API. And with Qwen4-Max, the company appears to be putting that jacket back on the hanger.
The released Qwen3.8-Max weights also represented a more limited version of the hosted product, lacking some of the capabilities available through Alibaba’s API, including native image input.
In other words, Alibaba was willing to wear the “open source jacket,” while keeping some of its most valuable capabilities behind a commercial API. And with Qwen4-Max, the company appears to be putting that jacket back on the hanger completely — no downloadable weights have been announced for the new model, and Alibaba has made no commitment to publish them.
On X and the accompanying blog post, Alibaba did say that the “open weights of smaller Qwen4 models will be released,” although it has yet to specify which models or when they will become available. But for Qwen4-Max itself, developers remain reliant on Alibaba’s hosted services or third-party API providers.
The post With Qwen4-Max, Alibaba puts the open source jacket back on the hanger appeared first on The New Stack.



