Alibaba just dropped its biggest Qwen yet. 2.4 trillion parameters. A million-token context window. Native vision.
Big numbers. But some of the details floating around are off. Let’s fix them.
Quick answer
Qwen4-Max is Alibaba’s new flagship AI model, launched November 16, 2026. It’s a Mixture-of-Experts model with 2.4 trillion total parameters and 71 billion active per token. It has a 1M-token context window and native vision. API pricing is $1.50 per million input tokens and $4.50 per million output tokens (international list). It’s closed-weight. The headline “~60% lower cost per task” is Alibaba’s own claim on its own benchmark.
What is Qwen4-Max?
It’s the top model in Alibaba’s Qwen family. Proprietary. You use it through the API on QwenCloud (Alibaba’s Model Studio), plus apps like QwenWork and QwenStudio.
Under the hood it’s a sparse Mixture-of-Experts. Think of a giant team where only a few experts work on each token. So the model is huge, but only 71 billion parameters fire per token. Alibaba says that’s 25% fewer active parameters than Qwen3.8-Max.
What are the key specs?
- Total parameters: 2.4 trillion
- Active per token: 71 billion
- Context window: 1 million tokens
- Vision: native. One review also lists video and PDF inputs.
- Tools: function calling, web search, code interpreter, MCP, structured output (per one review of the docs)
- Open weights: no. Smaller Qwen4 models are promised later, with no dates.
How much does Qwen4-Max cost?
Official international list pricing (Singapore):
- Input: $1.50 per million tokens
- Output (thinking plus answer): $4.50 per million tokens
- Cached input: $0.075 per million tokens
- Free quota: 1 million tokens (Singapore only, valid 90 days)
Heads up. Some summaries say $6.00 for output and 500,000 free tokens. That’s not what the pricing page shows. $6.00 is the output price of the older Qwen3.8-Max. So Qwen4-Max is cheaper per token than its predecessor ($2 in / $6 out).
Prices differ by region. Always check the live pricing page before you budget.
Does it really cut task costs by 62%?
Alibaba says it cuts average cost per task by “around 60%” on Work360, versus Qwen3.8-Max. It also says the model uses about 35% fewer tokens at high-effort reasoning. One review puts the Work360 numbers at $1.67 vs $4.43 per task. That’s where the 62% comes from.
Fine print:
- Work360 is Alibaba’s own benchmark. 721 specialized fields. Not independent.
- Artificial Analysis measured $3.27 per task on its own workload. That’s not comparable, but it shows costs vary by task.
- Your real savings depend on your prompts, not Alibaba’s chart.
How good is it compared to other models?
Alibaba’s chart says Qwen4-Max beats Qwen3.8-Max on all 15 benchmarks where both have scores. Examples from the chart: DeepSWE 75.7 vs 69.3, Terminal-Bench 4.0 48.5 vs 38.9, Humanity’s Last Exam 51.7 vs 43.9.
On Code Arena it ranks 6th (Elo 1750). Claude Opus 5.5 leads at 1816, per the same review.
Artificial Analysis lists an Intelligence Index score of 43. That’s the main independent figure so far. More outside tests should follow.
Is there a safety angle?
Yes. Alibaba says it filtered cybersecurity training data by intent, keeping defensive material and removing offensive attack content. It also added code-security checks during software training and extra CBRN safety work. These are Alibaba’s own statements.
What we still don’t know
- Independent results on real workloads.
- How it handles long 1M-token prompts in practice.
- When smaller open-weight Qwen4 models arrive.
- Whether the fourfold context jump claim holds. I couldn’t verify the previous model’s exact context size.
FAQ
When was Qwen4-Max released?
November 16, 2026, per Alibaba’s press release.
How many parameters does Qwen4-Max have?
2.4 trillion total, 71 billion active per token.
How much does Qwen4-Max cost?
$1.50 per million input tokens and $4.50 per million output tokens on the international list. Cached input is $0.075.
Is Qwen4-Max open source?
No. It’s closed-weight. Alibaba says smaller Qwen4 models will be open-weight later.
Does Qwen4-Max support vision?
Yes. It has native vision and a 1M-token context window.
Should I switch to Qwen4-Max?
Test it on your own tasks first. Use the free quota. Compare cost per finished task, not just per token.
Bottom line
Qwen4-Max is a serious flagship. Cheaper than its predecessor. Strong on Alibaba’s charts.
But most numbers are vendor-reported. Run your own test before moving production work.