Alibaba has released Qwen4-Max, the most capable model in its Qwen family with 2.4 trillion parameters and native vision.

The model also supports a context window of one million tokens.

Qwen4-Max ranked 6th on Code Arena, a real-world coding leaderboard. It also posted strong scores on benchmarks for agent workflows, workplace productivity and visual reasoning, such as Agents’ Last Exam and JobBench.

On Work360, Alibaba’s own benchmark covering 721 specialised fields, Qwen4-Max beats its predecessor Qwen3.8-Max overall. It uses 35 percent fewer tokens in High effort reasoning mode and cuts the average cost per task by around 60 percent.

The gains come from a new Qwen4 architecture that upgrades the sparse Mixture-of-Experts design across attention, residual connections and embeddings.

The model activates 71 billion of its 2.4 trillion parameters per token, which is 25 percent fewer than Qwen3.8-Max.

Alibaba has built cloud resource scheduling into the model. In testing, this cuts data-annotation model call costs by an average of nine percent and sped up offline CPU rescheduling more than 14 times.

In a car manufacturing demonstration, the model reduced manual workpiece programming time by 95 percent, down to 24 minutes from about 475 minutes. It caught overcuts and insufficient material removal before machining began.

In a science test, Qwen4 ran for 48 hours across four iterations of a closed-loop workflow to find an experimentally validated chip coolant molecule.

In an Alzheimer’s experiment it made 454 tool calls and 3,239 metric evaluations to design and optimise 24 candidate molecules.

Alibaba released Qwen3.8-Max on August 3 with the same 2.4 trillion parameters and context window. It introduced the QwenCloud platform at its Qwen Conference in Singapore on May 26.

Qwen4-Max is available now on QwenCloud through the Token Plan and API, and across QwenWork and QwenStudio. Input with an implicit cache hit costs US$0.075 per million tokens. Smaller open-weight Qwen4 models will follow soon.


Discover more from Entelechy Asia

Subscribe to get the latest posts sent to your email.

Leave a Reply