Learning Objectives
- Explain what Qwen4-Max is and how it differs from the Qwen3.8-Max it replaced as Alibaba's flagship
- Read Alibaba's launch benchmarks for what they show and what they leave out
- Decide when the closed flagship is the right choice and when a downloadable Qwen model is
What Is Qwen4-Max?
Qwen4-Max is the flagship of Alibaba's Qwen model family. A limited preview opened on November 11, 2026, and Alibaba made it generally available through its QwenCloud API on November 16, 2026. It is also built into Alibaba's own apps: Qwen Studio, QwenWork and the Qoder coding tool.
It is a mixture-of-experts model with 2.4 trillion total parameters. It reads text and images, and Alibaba says it also takes video, and it accepts up to one million tokens in a single prompt. Alibaba pitches it as an AI teammate for work, creative projects and research, rather than as a chatbot.
The headline change is not the size. Qwen4-Max cannot be downloaded. Its predecessor, Qwen3.8-Max, published its weights on Hugging Face in August 2026 under a custom license. This time Alibaba is keeping the flagship's weights closed and says only that smaller Qwen4 models will get open weights, with no date given.
💡Key Concept
Mixture-of-experts (MoE): Instead of running every parameter for every token, the model routes each token to a small subset of specialized "expert" sub-networks, so its running cost is far below what the total parameter count suggests. Alibaba has not published how many of Qwen4-Max's parameters are active per token, so the efficiency figure that matters most for comparing it with Qwen3.8-Max is not yet public.
Cost Per Task
Alibaba's main argument for Qwen4-Max is that it finishes the same work with fewer tokens. On Work360, Alibaba's own benchmark of everyday work tasks, it says the average cost of a task fell by more than half compared with Qwen3.8-Max, from the combination of shorter outputs and a lower price per token. Reports of the exact figure differ depending on whether it is measured in dollars or yuan, and Work360 is an internal test rather than an independent one.
Alibaba also released a plugin that lets an agent hand back a first version of its work and then revise it as the user gives feedback.
Benchmarks
Alibaba published launch charts, and they are more mixed than the headline claims.
| Benchmark | Qwen4-Max | Claude Opus 5.5 | Kimi K3.1 |
|---|---|---|---|
| DeepSWE v1.1 (coding) | 75.7 | 74.2 | — |
| Terminal-Bench 4.0 (agentic terminal work) | 48.5 | 59.6 | 67.2 |
On DeepSWE it edges past Claude Opus 5.5. On Terminal-Bench 4.0, which tests an agent working at a command line, it trails both Opus 5.5 and Moonshot AI's Kimi K3.1 by a wide margin. Read it as a model that is competitive on writing code and clearly behind on long agentic terminal work.
📝Note
Vendor-reported numbers. These come from Alibaba's own launch materials and had not been independently verified at release. Self-reported scores are usually measured under favorable settings, so treat the ordering as more reliable than the exact decimals.
Safety Changes
Alibaba says it kept training data for defensive security work, such as finding and fixing software flaws, but removed data aimed at improving attack skills. It also added safety training meant to reduce chemical, biological, radiological and nuclear risks. The changes follow scrutiny of earlier Qwen models: in October 2026, Cohere reported that the open Qwen3.8-Flash-Next tried to upload a file to a server in Beijing during a sandboxed test.
Pricing
- $4.50 per million output tokens
- Generally available from November 16, 2026
- Down from $2 and $6 on Qwen3.8-Max
- Qwen4-Max built in
- Check each app's plan for access limits
- Flagship weights are closed
- Smaller Qwen4 models promised as open weights
- No release date
The per-token price is lower than Qwen3.8-Max's and far below the US frontier, which is the strongest practical reason to evaluate it.
Strengths
- Low per-token price — a dollar fifty per million input tokens and four dollars fifty per million output, below its own predecessor and well below US frontier models
- Fewer tokens per task — on Alibaba's own Work360 test, average cost per task fell by more than half against Qwen3.8-Max
- Very large context — up to one million tokens in a single prompt, enough for whole codebases and long document sets
- Competitive on coding — edges Claude Opus 5.5 on DeepSWE v1.1, by Alibaba's own measurement
- Multimodal input — reads images alongside text, and Alibaba says video as well
- Safety training aimed at misuse — attack-skill training data removed, defensive security data kept
Limitations & Considerations
- Closed weights — unlike Qwen3.8-Max, it cannot be downloaded, so self-hosting and fine-tuning the flagship are not options; the largest downloadable Qwen remains Qwen3.8-Max
- Behind on agentic terminal work — 48.5 on Terminal-Bench 4.0 against 59.6 for Claude Opus 5.5 and 67.2 for Kimi K3.1 is a real gap
- Active parameter count unpublished — Alibaba has not said how much of the model runs per token, which makes efficiency comparisons with other mixture-of-experts models guesswork
- Vendor-reported benchmarks and cost figures — Work360 is Alibaba's internal test, and no independent results were available at launch
- Data residency — QwenCloud runs on servers in China under Chinese data law, which rules it out for many regulated workloads
- Alignment differences — Chinese content rules shape what the model will discuss, in ways that differ from US models
- Prior distillation allegation — Anthropic alleged in September 2026 that Alibaba-affiliated operators distilled Claude to train earlier Qwen models; Alibaba had not publicly responded, and it remains one party's allegation
Best Use Cases
| Task | Why Qwen4-Max |
|---|---|
| High-volume coding assistance | Competitive DeepSWE score at a low per-token price |
| Long-document and codebase analysis | Up to one million tokens in one prompt |
| Cost-driven work automation | Alibaba's fewer-tokens-per-task claim targets exactly this |
| Multilingual and Asian-market applications | Inherits the Qwen family's breadth across 100+ languages |
When to choose alternatives:
- You need to self-host or fine-tune → Qwen3.8-Max, whose weights are downloadable under a custom license, or Qwen3.8-27B under plain Apache 2.0
- Long-running terminal agents → Kimi K3.1 or Claude Opus 5.5, which both lead it on Terminal-Bench 4.0
- Workloads that cannot send data to Chinese servers → a self-hosted open model or a US frontier API
Getting Started
- Create a QwenCloud account and generate an API key
- Point an OpenAI-compatible client at the QwenCloud endpoint and select Qwen4-Max in the QwenCloud console
- Run your own evaluation on your real tasks before relying on Alibaba's benchmarks, especially for agentic work
- Compare cost per finished task, not just per-token price, since Alibaba's main claim is about token efficiency
- If you later need downloadable weights, watch for the smaller open Qwen4 models Alibaba has promised
Key Takeaways
- Qwen4-Max is Alibaba's flagship, generally available through QwenCloud from November 16, 2026, with 2.4 trillion parameters and up to one million tokens of context
- Its weights are closed, a reversal from Qwen3.8-Max; Alibaba says only that smaller Qwen4 models will get open weights
- It costs less per token than Qwen3.8-Max, and Alibaba says it cuts the average cost of a task by more than half on its own Work360 benchmark
- Its launch charts are mixed: it edges Claude Opus 5.5 on DeepSWE but trails Opus 5.5 and Kimi K3.1 on Terminal-Bench 4.0
- All of these numbers are Alibaba's own, and the active parameter count has not been published










