
Alibaba has released Qwen4-Max, and the pitch is less about being the biggest model and more about doing more work per token. Announced on November 16, 2026, the new flagship of the Qwen family has 2.4 trillion total parameters, native vision and a one-million-token context window, yet it activates only 71 billion parameters for each token it processes.
What Alibaba released
Per Alibaba Cloud’s press release, Qwen4-Max is the most capable model in the Qwen family so far. It is available now on QwenCloud through a Token Plan and API, and across Alibaba platforms including QwenWork and QwenStudio. Input with an implicit cache hit is priced at $0.075 per million tokens. Alibaba also says smaller open-weight models from the Qwen4 series will be released to developers soon, though it gave no date.
A sparser architecture
The model uses a sparse Mixture-of-Experts design, where only a slice of the network runs for any given token. Alibaba says the new Qwen4 architecture upgrades the design across attention, residual connections and embeddings. Compared with Qwen3.8-Max at the same 2.4 trillion scale, Qwen4-Max activates 25 percent fewer parameters per token, which the company links to stronger capability at lower inference cost.
Benchmarks and efficiency claims
Alibaba reports that Qwen4-Max ranked 6th on Code Arena, a real-world coding leaderboard, and posted strong results on agent, workplace and visual reasoning tests including Agents’ Last Exam, Vals Index, JobBench and Chartography. On Work360, a proprietary benchmark covering 721 specialized fields, the company says Qwen4-Max beats its predecessor overall while using 35 percent fewer tokens in High effort reasoning mode and cutting average cost per task by around 60 percent. These are vendor-reported results, and Work360 is Alibaba’s own test, so independent evaluations will be the real check.
Agents that run for hours
The release leans heavily on long-running agent work. Alibaba describes several showcases:
- Cloud scheduling: in a closed loop, the model optimized its own scheduling workflows, cutting data-annotation model call costs by an average of 9 percent and speeding offline CPU rescheduling by more than 14 times.
- Manufacturing: in an automotive showcase, it cut manual workpiece programming time by 95 percent, from about 475 minutes to 24, by iterating on feedback to catch overcuts and insufficient material removal before machining.
- Science: in a continuous 48-hour run over four iterations of design, synthesis and testing, and active learning, it helped discover an experimentally validated coolant molecule for next-generation chip cooling.
- Drug research experiment: in an Alzheimer’s research experiment it made 454 tool calls and 3,239 metric evaluations to design and optimize 24 candidate molecules, all computational.
The model is also pitched for 3D modeling, game development and agentic video creation, and for forward-deployed engineering in finance, law and healthcare.
Safety by design
Alibaba says governance runs through the model’s lifecycle. Software engineering training paired code security checks with functionality tests, and training data was filtered by task intent so defensive material was kept and offensive attack content removed. Dedicated safety datasets and post-training aim to strengthen protections against chemical, biological, radiological and nuclear risks.
What it means
Qwen4-Max shows the frontier race shifting toward efficiency: fewer active parameters, fewer tokens, lower cost per task. For teams building agents, a cheaper long-horizon model changes which jobs are worth automating. The caution is that the headline numbers come from Alibaba’s own benchmarks and showcases, so it is worth testing on your own workloads once access is available.
Conclusion
With a one-million-token window, native vision and a lean per-token footprint, Qwen4-Max is aimed squarely at sustained agentic work. The promised open-weight Qwen4 models could matter just as much for developers.