The first Qwen4 model has native vision and a one-million-token window, and Alibaba says it costs about 60 percent less per task than Qwen3.8-Max.
· 3 min read
Alibaba released Qwen4-Max on November 16, the first model in its Qwen4 family and, by the company's own account, the most capable Qwen model to date. It is a 2.4-trillion-parameter mixture-of-experts system with native vision and a one-million-token context window, and it is offered through Alibaba's cloud rather than as a download.
What Alibaba says it can do
According to Alibaba Cloud's announcement, the new architecture activates 71 billion parameters per token, which is 25 percent fewer than Qwen3.8-Max at the same 2.4-trillion scale. Alibaba ranks the model sixth on Code Arena, a real-world coding leaderboard, and cites strong scores on Agents' Last Exam, Vals Index, JobBench and Chartography.
The headline efficiency claim rests on Work360, a proprietary Alibaba benchmark covering challenging tasks across 721 specialised fields. On it, Qwen4-Max beats Qwen3.8-Max overall while using 35 percent fewer tokens at its High effort setting and cutting the average cost per task by around 60 percent. DataCamp's breakdown of the launch puts the Work360 scores at 52.69 for Qwen4-Max and 47.74 for its predecessor.
Independent results are more mixed
Because Work360 is Alibaba's own test, outside numbers matter. AF.net reports that Alibaba's launch research has Qwen4-Max scoring 75.7 on the DeepSWE v1.1 software engineering benchmark and leading Claude Opus 5.5 there, while independent evaluations cited by the same outlet place it behind rivals such as Kimi, though third-party tests still back its cost efficiency per task. That split is common at a launch, and it means the strongest claims will need testing by people outside Alibaba.
Pricing and access
Qwen4-Max is available now on QwenCloud through the Token Plan and the API, and across Alibaba platforms including QwenWork and QwenStudio. Alibaba lists input at a cache hit at $0.075 per million tokens. Trackers such as DataCamp and AI-TLDR put the standard price at $1.50 per million input tokens and $4.50 per million output tokens, which is below Qwen3.8-Max. A preview opened on November 11 for Token Plan users only and, per AI-TLDR, is due to be retired after November 25.
The model itself will not be downloadable. The New Stack describes Qwen4-Max as API-only and notes that it follows Alibaba's decision to open Qwen3.8-Max earlier this year, which makes this release a return to the older closed approach for the Max line. Alibaba says smaller, open-weight Qwen4 models will soon be released to developers worldwide, but it has not given dates or a list of sizes.
Agent demonstrations
Alibaba built much of the launch around long-running agent work, and TechNode adds that the company also released a plugin that lets agents deliver an initial version of their work and revise it as users give feedback. In a cloud resource scheduling loop, the model reduced data-annotation model call costs by an average of 9 percent and sped up offline CPU rescheduling by more than 14 times. In an automotive manufacturing showcase, it cut manual workpiece programming time by 95 percent, from roughly 475 minutes to 24, by catching overcuts and insufficient material removal before machining began.
In science, Alibaba says Qwen4 completed a continuous 48-hour agent run through four iterations of a closed-loop workflow, combining molecular design, synthesis and testing, and active learning, to find an experimentally validated coolant molecule for chip cooling. In an Alzheimer's research experiment, the model executed 454 tool calls and 3,239 metric evaluations to design and optimise 24 candidate molecules. These are company showcases, not independent replications.
Safety measures
Alibaba says it embedded governance across the model's lifecycle. Software engineering training included code security checks alongside functionality tests, which it says significantly reduced vulnerabilities such as injection, memory safety and broken access control. For cybersecurity, it filtered training data by task intent, keeping defensive material and removing offensive attack content, and it added dedicated safety datasets to strengthen protections against chemical, biological, radiological and nuclear risks.
Why it matters
The release shows Chinese labs competing on cost per task as much as on raw scores, at a time when open-weight models from Meta and Mistral are also pushing the field. We covered Meta's Muse Spark Open and Mistral's Large 4 open weights earlier, and Qwen models already sit behind some features in Apple Intelligence testing in China. For developers who depend on open weights, the news here is the promise of smaller Qwen4 releases, not Qwen4-Max itself.
The open questions are practical: how the model performs on independent coding and agent tests, whether the lower price holds as demand grows, and when the open-weight Qwen4 models actually arrive.
Cover image: Alibaba Cloud.
Published November 18, 2026. Gadgets365 will update this article as more information becomes available.
Editor
Pravin KankanePravin Kankane is the editor of Gadgets365, an independent technology publication based in Pune, India. He covers smartphones, AI, Apple, Android and consumer tech launches in India.