Alibaba Keeps Qwen4-Max Closed, Claims 60% Lower Task Costs Days Before Nov. 19 Results
The new flagship model is sold only through Alibaba's own cloud and apps, a break with the open releases that made Qwen popular with developers, three days before the company reports quarterly results.

What to know
- Alibaba released Qwen4-Max on Monday, a mixture-of-experts model with 2.4 trillion total parameters that activates 71 billion per token.
- The company said the model beats Qwen3.8-Max on its Work360 benchmark while cutting the average cost per task by about 60%.
- Alibaba is offering the flagship only through QwenCloud and its own apps, with open weights promised for smaller Qwen4 models; it reports September-quarter results on Nov. 19.
HONG KONG
Alibaba Group released Qwen4-Max on Monday, a 2.4 trillion-parameter artificial intelligence model that the company says completes complex work tasks at roughly 60% lower average cost than its predecessor. Alibaba is offering it only through its own cloud service and apps, and has promised open weights for smaller Qwen4 models, a break with many earlier Qwen releases that The Next Web highlighted.
The launch lands three days before Alibaba reports results for the September quarter on Nov. 19, according to a board-meeting notice the company filed with the U.S. Securities and Exchange Commission. For investors in Hong Kong, a central question is whether the company can turn years of free, downloadable models into paid cloud revenue. A closed flagship is a direct bet on that.
Qwen4-Max is available through QwenCloud, Alibaba's model service, on its Token Plan subscription and through the pay-as-you-go application programming interface, as well as in the company's QwenWork and QwenStudio apps, Alibaba said in a statement on its Alizila news site. Business Standard reported the same availability. Smaller Qwen4 models will be released with open weights to developers worldwide, Alibaba said, without giving a date.
What Alibaba says Qwen4-Max does
The model uses a sparse mixture-of-experts design, which routes each request through only part of the network. Of its 2.4 trillion total parameters, 71 billion are active for each token processed, 25% fewer than Qwen3.8-Max at the same overall size, according to the company. It reads text, images and video and accepts prompts of up to one million tokens.
On Work360, an internal Alibaba benchmark covering tasks in 721 specialized fields, Qwen4-Max scored better than Qwen3.8-Max overall while using 35% fewer tokens in its high-effort reasoning mode, the company said. Alibaba put the reduction in average cost per task at about 60%. The Next Web, citing Alibaba's figures, reported that the average cost on that test fell from $4.43 to $1.67.
The company said the model ranked sixth on Code Arena, a public coding leaderboard, and published results on agent and workplace benchmarks including Agents' Last Exam, Vals Index and JobBench. The Next Web said it had not verified Alibaba's benchmark claims and noted that charts released with the model gave a mixed picture: on Terminal-Bench 4.0, Qwen4-Max scored 48.5 against 67.2 for Moonshot AI's Kimi K3.1, according to figures it attributed to FourWeekMBA.
Alibaba also pointed to industrial demonstrations. In one automotive showcase, it said, the model cut manual workpiece programming time by 95%, from about 475 minutes to 24 minutes. In a chip-cooling project, Qwen4 ran continuously for 48 hours to find a coolant molecule that was later validated by experiment, the company said. Business Standard, citing Reuters, reported that Alibaba's internal safety testing found no evidence of significant high-risk capabilities.
Pricing and the closed-weights shift
The Next Web reported API prices of $1.50 per million input tokens and $4.50 per million output tokens. Alibaba's own statement listed input that hits the implicit cache at $0.075 per million tokens, a discount aimed at repeated prompts such as long system instructions. QwenCloud's developer documentation says the model can be called through OpenAI-style interfaces as well as Alibaba's own DashScope format, and lists setup guides for popular coding tools.
The decision on weights matters because openness built Qwen's reach. According to Bloomberg, as cited by The Next Web, Qwen is already the most downloaded open-weight model family in the world. Developers could run earlier flagships on their own hardware without paying Alibaba. With Qwen4-Max, the only way in is through Alibaba's cloud or apps. A limited preview opened on Nov. 11, The Next Web reported, and QwenCloud's documentation lists a separate preview model alongside the full release.
The apps are part of the funnel. Alibaba said in October that QwenWork, its workplace agent platform launched in August, drew more than 30 million users in its first month, over half of them enterprise users. Capital Times reported earlier on AI token packages selling during Singles' Day and on Alibaba's plan to list its T-Head chip unit.
Tencent first, then Alibaba
China's largest internet companies report back to back this week. Tencent's board meets on Wednesday, Nov. 18, to approve its third-quarter results, according to the company's exchange notice. Alibaba's board will review its results for the three and six months to Sept. 30 on Nov. 18, with the announcement set for Nov. 19, its filing shows.
Investors will look in both reports for the size of AI-related capital spending and whether cloud revenue is growing fast enough to pay for it. For Alibaba, Thursday's figures will be the first read on cloud demand since the company began steering its best model toward paid access, although Qwen4-Max arrived too late to show up in September-quarter revenue. The Qwen4 open-weight models, which Alibaba has promised but not dated, are the next release to watch.
Sources
- Alizila (Alibaba Group): Alibaba Unveils Qwen4-Max: New AI Teammate for Work, Creativity, and Discovery (Nov. 16, 2026)
- Business Standard: Alibaba says Qwen4-Max cuts task costs by 60%, boosts cyber safeguards (Nov. 16, 2026)
- The Next Web: Alibaba launches Qwen4-Max flagship AI model without open weights (Nov. 16, 2026)
- QwenCloud documentation (Alibaba): Latest model: Qwen4-Max (Nov. 16, 2026)
- Alizila (Alibaba Group): QwenWork Unveiled E-Commerce Tool To Help Merchants in China Boost Efficiency (Oct. 14, 2026)
- Alibaba Group via U.S. SEC: Date of Board Meeting (Alibaba Group Holding Ltd.), Form 6-K Exhibit 99.1 (Nov. 6, 2026)
- Tencent via HKEXnews: Date of Board Meeting (Tencent Holdings Limited) (Sept. 24, 2026)
Our standards: Capital Times reports under the principles set out in our Editorial Standards and Ethics Policy. Spotted an error? Report it to our editors.







