Alibaba has released Qwen4-Max, the newest model in its Qwen family, built around a sparse mixture-of-experts design that keeps running costs down even as the model grows. The launch comes from Alibaba‘s AI unit and puts the model up against rival frontier systems on cost as much as capability.
Key Takeaways
- Qwen4-Max activates just 71 billion of its 2.4 trillion parameters per token, 25% fewer than predecessor Qwen3.8-Max at the same scale, while cutting average task cost by around 60% on Alibaba’s Work360 benchmark.
- In an automotive manufacturing demonstration, the model cut manual workpiece programming time by 95%, from roughly 475 minutes to 24 minutes.
- Qwen4-Max ran a continuous 48-hour agentic workflow that discovered an experimentally validated coolant molecule for next-generation chip cooling.
A Model Built to Use Fewer Tokens
Qwen4-Max has 2.4 trillion total parameters, but activates just 71 billion of them per token — 25 per cent fewer activated parameters than its predecessor, Qwen3.8-Max, at the same 2.4 trillion scale. That selective activation is the core of the sparse mixture-of-experts approach, where only a subset of the model’s “experts” switches on for each piece of input.
The new Qwen4 architecture extends that sparse design further, upgrading how the model handles attention, residual connections and embeddings. On Work360, a proprietary benchmark that tests performance across 721 specialised industry fields, Qwen4-Max beats Qwen3.8-Max overall while using 35 per cent fewer tokens in its high-effort reasoning mode — cutting the average cost per task by around 60 per cent.
Vision, Context and Benchmark Scores
Qwen4-Max adds native vision capabilities and a one-million-token context window, letting it process far longer documents or conversations in a single pass. On Code Arena, a leaderboard that ranks models on real-world coding tasks, it placed sixth. It also posted strong results on Agents’ Last Exam, the Vals Index, JobBench and Chartography, benchmarks that test agentic behaviour, workplace productivity and visual reasoning respectively.
Cutting Costs From Cloud Infrastructure to the Factory Floor
Qwen4-Max also includes Cloud Resource Scheduling, which lets the model optimise its own scheduling workflows in a closed loop. In Alibaba’s benchmark testing, this cut data-annotation model call costs by an average of 9 per cent and sped up offline CPU rescheduling by more than 14 times.
The model’s reach extends to physical production as well. In an automotive manufacturing demonstration spanning production-line planning, part design and machining verification, Qwen4-Max cut manual workpiece programming time by 95 per cent — from roughly 475 minutes down to 24 minutes — by using feedback-driven iteration to catch overcuts and insufficient material removal before machining begins. Alibaba says the model also speeds up forward-deployed engineering work in finance, law and healthcare.
Long-Running Agents for Scientific Discovery
Qwen4-Max is built to sustain long agent runs rather than respond to single prompts. To tackle next-generation chip cooling, it completed a continuous 48-hour run through four iterations of a closed-loop workflow — combining molecular design, synthesis and testing, and active learning — that ended in an experimentally validated coolant molecule.
In a separate Alzheimer’s research exercise, Qwen4-Max worked through target research, molecular design and computational evaluation, autonomously making 454 tool calls and 3,239 metric evaluations to design and optimise 24 candidate molecules.
Safety Checks Built Into Training
Alibaba says it has built governance into the model across its lifecycle. Software engineering training included code security checks alongside functionality tests, aimed at cutting vulnerabilities such as injection flaws, memory-safety bugs and broken access control. For cybersecurity specifically, training data was filtered by intent, keeping material useful for vulnerability detection and remediation while stripping out offensive attack content. Alibaba also added dedicated safety datasets and post-training work aimed at limiting chemical, biological, radiological and nuclear (CBRN) risks.
Where Enterprises Can Access It
Qwen4-Max is available now on QwenCloud through a Token Plan and API, as well as through QwenWork and QwenStudio. Input is priced at $0.075 per million tokens when the implicit cache is hit. Alibaba says smaller, open-weight models from the Qwen4 series will follow for developers worldwide.



