Alibaba released Qwen4-Max on November 16, 2026, describing it as the most capable model in the Qwen family so far. The model has 2.4 trillion total parameters, activates 71 billion for each token it processes, can handle images natively and has a context window of up to one million tokens. A context window is the amount of text and other input a model can consider at once.
People can use Qwen4-Max through Alibaba Cloud’s QwenCloud Token Plan and API, and through Alibaba products QwenWork and QwenStudio. The release is aimed at tasks ranging from coding and workplace jobs to visual reasoning and research. Alibaba says the model ranked sixth on Code Arena, a coding leaderboard, and performed well on several other benchmarks. Those are company-reported results; the announcement does not give detailed scores for each test.
How we got here
Qwen4-Max uses a sparse Mixture-of-Experts design, which routes each request through only part of a model rather than using all its parameters at once. Alibaba says its new architecture changes attention, residual connections and embeddings. At 71 billion active parameters, it activates 25% fewer parameters per token than Qwen3.8-Max, its previous model at the same 2.4-trillion-parameter scale. Alibaba says this makes Qwen4-Max more capable while reducing the cost of running it.
For a direct comparison, Alibaba used Work360, its own benchmark covering challenging tasks across 721 specialized fields. It says Qwen4-Max did better overall than Qwen3.8-Max while using 35% fewer tokens in High effort reasoning mode and lowering average cost per task by about 60%. MarkTechPost reports a 62% cost reduction. The figures are close, but the sources do not explain the difference; Work360 is proprietary, so readers cannot independently inspect its tasks or results from these details.
What Alibaba says it can do
Alibaba also describes uses beyond chat and coding. In a cloud-infrastructure test, it says Qwen4-Max reduced the cost of data-annotation model calls by an average of 9% and made offline CPU rescheduling more than 14 times faster. In an automotive manufacturing demonstration, the company says the model cut manual workpiece programming from roughly 475 minutes to 24 by iterating on designs and checking for machining errors. These are examples and measurements reported by Alibaba, not independent evaluations.
The company describes two science projects involving long-running AI agents—systems that can carry out a sequence of tasks using tools. In one, Qwen4 ran continuously for 48 hours and completed four rounds of molecular design, synthesis, testing and learning to identify a coolant molecule that was experimentally validated. In an Alzheimer’s research experiment, Alibaba says Qwen4-Max made 454 tool calls and 3,239 evaluations while designing and refining 24 candidate molecules. The announcement does not establish that those candidates are treatments or that the work has been independently assessed.
What happens next
Alibaba lists Qwen4-Max’s input price as $0.075 per million tokens for an “Implicit Cache Hit.” MarkTechPost gives $1.50 per million input tokens, without specifying the same cache condition. These are different figures for differently described pricing, so the lower cached-input price should not be read as the general cost of every request. The available details do not provide a full price comparison across input, output and cache conditions.
Alibaba says smaller, open-weight models from the Qwen4 series will be released to developers worldwide soon, but gives no date. Open-weight models make their model weights available for others to use or adapt. For now, Qwen4-Max is available via Alibaba’s cloud services and apps; the announcement does not say that the flagship model itself is open-weight.
Why it matters
For developers and companies already using Alibaba Cloud, Qwen4-Max is available through its API and Token Plan, so they can try it without waiting for the smaller open-weight models. Alibaba’s Work360 results suggest some business tasks may cost less than with Qwen3.8-Max, but the benchmark is proprietary and the reported prices describe different input conditions. The actual cost for a particular workload—and whether Qwen4-Max is cheaper than alternatives—remains unclear from the published details.