Alibaba releases Qwen4-Max, a 2.4 trillion-parameter model that activates 71 billion, and prints one cache-hit price
The Hangzhou release is dated 16 November. The Model Studio page adds a qwen4-max entry that day and still describes the 11 November preview as Token Plan only. The only dollar figure in the release is $0.075 per million tokens for an implicit cache hit. TechNode’s task-cost cut is about 58 percent. The release says around 60.
Feature3 minIn-depth
James Okonkwo3 min read
- Qwen
- Alibaba
- Agents
- API

Photo: TechNodeSource page
Alibaba on Monday released Qwen4-Max, which its press note, datelined Hangzhou and dated 16 November, calls the most capable model in the Qwen family. The note gives it 2.4 trillion total parameters, native vision, and a one-million-token context window. It is a sparse mixture-of-experts model that activates 71 billion parameters per token, which Alibaba says is 25 percent fewer activated parameters than Qwen3.8-Max at the same 2.4 trillion scale. It is available now on QwenCloud through the Token Plan and the API, and on QwenWork and QwenStudio. The Model Studio releases page adds a 16 November entry for qwen4-max, international deployment, and leaves the 11 November entry for qwen4-max-preview in place, still described as a preview available through Token Plan only.
The only dollar price in the press note is a partial one. “Input (Implicit Cache Hit) is priced at $0.075 per 1M tokens.” The note does not print a standard input rate, an output rate, or a cache-miss price. On its own Work360 benchmark, which it describes as challenging industry tasks across 721 specialized fields, Alibaba says Qwen4-Max outperforms Qwen3.8-Max overall while using 35 percent fewer tokens at High effort and reducing the average cost per task by around 60 percent. TechNode’s story the same day prints figures the release does not: average task cost on that benchmark down from RMB 26.58 to RMB 11.16, a reduction of about 58 percent. The two percentages are not the same number. TechNode also says Alibaba released a plugin that lets agents deliver a first version of work and revise it as users respond, and that the model is available in Qoder as well as Qwen Studio and QwenWork. The press note does not mention the plugin or Qoder.
A sixth-place coding rank, and scores the note does not print
The ranking Alibaba does print is sixth on Code Arena, which it calls a real-world coding leaderboard. It says the model posted strong scores on Agents’ Last Exam, the Vals Index, JobBench, and Chartography, and it does not attach a number to any of those four. The architecture line is that Qwen4 upgrades the sparse mixture-of-experts design across attention, residual connections, and embeddings. Smaller Qwen4 models “will soon be released” with open weights. TechNode’s wording is that Alibaba plans to release smaller Qwen4 models as open source. The company has not given a date. The flagship in this note is not described as downloadable.
The applied claims are also the company’s, and they are not all pinned to the Max name. In a cloud-scheduling test, Alibaba says the model cut data-annotation call costs by an average of 9 percent and sped offline CPU rescheduling by more than 14 times. In an automotive showcase, it says manual workpiece programming fell by 95 percent, from roughly 475 minutes to 24 minutes, because the model caught overcuts and insufficient material removal before machining. On chip cooling, the note says “Qwen4,” not Qwen4-Max, completed a continuous 48-hour agent run through four iterations of molecular design, synthesis and testing, and active learning, and discovered an experimentally validated coolant molecule. In an Alzheimer’s experiment it does attribute to Qwen4-Max, the model made 454 tool calls and 3,239 metric evaluations and designed 24 candidate molecules. Those are showcase and experiment counts in a press note. They are not an independent paper.
The safety paragraph says software-engineering training included code-security checks alongside functionality tests, and that cybersecurity training data was filtered by task intent, keeping defensive material for finding and fixing vulnerabilities and removing offensive attack content. It also says dedicated safety datasets and post-training work are meant to strengthen protections against chemical, biological, radiological, and nuclear risks. The note does not publish a score for any of those filters. What a developer can check on Monday is narrower: a named model on the API and the Token Plan, one cache-hit price, a parameter split of 2.4 trillion and 71 billion, and a preview entry that the releases page has not yet taken down.


