Skip to content
AI.info

The Pulse

Alibaba releases Qwen4-Max with a million-token context window

Alibaba has released Qwen4-Max, a multimodal model with 2.4 trillion total parameters and native vision capabilities. The company says it is available through QwenCloud and plans smaller open-weight Qwen4 models.

Alibaba releases Qwen4-Max with a million-token context window

Alibaba’s new Qwen4-Max model has a one-million-token context window and native vision capabilities, the company said in a November 16 release. The model has 2.4 trillion total parameters and is the latest flagship in Alibaba’s Qwen family.

Qwen4-Max is available on QwenCloud through the Token Plan and API. It can also be accessed through Alibaba platforms including QwenWork and QwenStudio, while smaller open-weight models in the Qwen4 series are planned for release to developers worldwide.

Architecture and access

The model uses a sparse mixture-of-experts architecture. Alibaba says it activates 71 billion parameters per token, 25% fewer than Qwen3.8-Max, which has the same 2.4 trillion total parameter count. The architecture updates the mixture-of-experts design across attention, residual connections and embeddings.

Alibaba Cloud’s documentation lists text generation, image and video understanding, web search, tool calling and structured outputs among the model’s capabilities. It also lists compatibility with the OpenAI and Anthropic APIs.

The release lists a price of $0.075 per 1M tokens for input under an implicit cache hit. A separate Qwen4-Max-Preview remains available only through the Token Plan; its preview period runs from 10:00 on November 11 to 24:00 on November 25, 2026, UTC+8, after which it will be retired.

Benchmark claims

On Alibaba’s proprietary Work360 benchmark, which evaluates tasks across 721 specialized fields, Qwen4-Max outperformed Qwen3.8-Max overall, the company says. It used 35% fewer tokens in High effort reasoning mode and reduced the average cost per task by around 60%, according to the release.

Alibaba also says Qwen4-Max ranked sixth on Code Arena, a coding leaderboard, and reports results on tests for agent workflows, workplace productivity and visual reasoning. In cloud resource scheduling benchmarks, the company says the model reduced data-annotation model call costs by an average of 9% and accelerated offline CPU rescheduling speeds by more than 14 times.

In a manufacturing showcase, Alibaba says the model cut manual workpiece programming time by 95%, from roughly 475 minutes to 24 minutes. The demonstration used feedback-driven iteration to identify overcuts and insufficient material removal before machining.

Research tasks and safeguards

Alibaba describes a 48-hour agent run that went through four iterations of molecular design, synthesis and testing, and active learning to discover an experimentally validated coolant molecule for chip cooling. In a separate Alzheimer’s research experiment, the company says Qwen4-Max made 454 tool calls and 3,239 metric evaluations to design and optimize 24 candidate molecules.

For software engineering training, Alibaba says it included code security checks alongside functionality tests. The company says it also filtered cybersecurity training data by task intent, retaining defensive material for vulnerability detection and remediation while removing offensive attack content.

Dedicated safety datasets and post-training changes were used to strengthen protections against chemical, biological, radiological and nuclear risks, according to Alibaba.

Sources

  1. Alibaba Cloud
  2. Alibaba Cloud
  3. Alibaba Cloud

Related stories

Explore

More articles