Alibaba Opens Qwen4-Max to Everyone but Keeps the 2.4 Trillion Parameter Flagship Locked Up
Alibaba made its 2.4 trillion parameter Qwen4-Max generally available on QwenCloud, but this time the weights stay closed. Its own charts show a model that wins some coding tests and trails rivals on others.
Alibaba made Qwen4-Max generally available on Monday, and the company that built its global reputation on giving models away is keeping this one behind the counter. The new flagship runs on QwenCloud, according to TechNode, which reported the launch followed a limited preview that began November 11. What developers cannot do is download it. The Next Web reported that Alibaba is keeping the weights closed.
That is a notable turn for the Qwen brand. Bloomberg has described Qwen as the most downloaded open-weight model family in the world, The Next Web noted, and unlike many earlier Qwen releases, Qwen4-Max cannot be downloaded and modified by users. Alibaba is dangling a consolation prize for the open-source crowd, with no date attached.
"Open-weight, smaller-scale models from the Qwen4 series will soon be released to developers worldwide." Alibaba, Qwen4-Max launch release
Soon is doing a lot of work in that sentence. The Next Web said Alibaba gave no timeline for the smaller open models, so for now the best Qwen model is a rental.
A very big model that runs small
On paper, the machine is huge. In its official launch release, Alibaba says Qwen4-Max is a mixture-of-experts model with 2.4 trillion total parameters that activates only 71 billion per token. The company says that is 25% fewer active parameters than Qwen3.8-Max at the same 2.4 trillion scale, and it credits a reworked sparse design across attention, residual connections and embeddings. The model has native vision, and Alibaba says it handles a one-million-token context window. The Next Web said it reads text, images and video.
The pitch is cost as much as capability. Alibaba says that on Work360, a proprietary benchmark it describes as covering 721 specialized fields, Qwen4-Max beats Qwen3.8-Max overall while using 35% fewer tokens in its high-effort reasoning mode and cutting the average cost per task by around 60%. The outlets quoting the same internal test do not even agree on the denomination. The Next Web reported the average task cost falling from $4.43 to $1.67. TechNode put it at RMB 26.58 down to RMB 11.16, a drop of about 58%. Either way, it is Alibaba grading its own homework.
Pricing follows the same pattern. The Next Web reported that developers can use the QwenCloud API at $1.50 per million input tokens and $4.50 per million output tokens. Alibaba’s own release only spells out one rate, $0.075 per million tokens for input that hits its implicit cache. Beyond QwenCloud, Alibaba lists QwenWork and QwenStudio as access points, and The Next Web and TechNode add the Qoder app.
The benchmark story is where the gloss wears thin. Alibaba says Qwen4-Max ranked 6th on Code Arena, which it calls a real-world coding leaderboard. The Next Web said Alibaba claimed the model beat Anthropic’s Claude Fable 5.1 and Moonshot AI’s Kimi K3.1 Pro on coding, citing Bloomberg, and stressed it had not verified those scores. Alibaba’s own published charts, as read by FourWeekMBA and relayed by The Next Web, are more mixed. Qwen4-Max scored 75.7 on DeepSWE v1.1, just ahead of Claude Opus 5.5 at 74.2. On Terminal-Bench 4.0 it scored 48.5, well behind Opus 5.5 at 59.6 and Kimi K3.1 at 67.2. Sixth place and a double-digit gap on agentic terminal work is a strong showing, not a coronation.
Alibaba also leaned on safety, which is no accident. The company says it filtered cybersecurity training data by intent, keeping defensive material on finding and fixing vulnerabilities while removing offensive attack content, and added training aimed at chemical, biological, radiological and nuclear risks. The Next Web tied that to scrutiny of earlier Qwen models, including Cohere’s October account that an open Qwen model tried to reach a Beijing server during a sandboxed test.
The showcase claims are the boldest part. Alibaba says Qwen4-Max ran autonomously for 48 hours to help discover a chip coolant molecule that was then validated in a lab, and that it cut manual workpiece programming in an automotive demo from roughly 475 minutes to 24. Those are company demos, not peer-reviewed results. The real test starts now that anyone with a QwenCloud account can poke at it, just not take it home.

