AI Models

Qwen4-Max

Checked November 20, 2026 on Alibaba Cloud Model Studio. International (Singapore) endpoint: $1.50 per million input tokens and $4.50 per million output, the same rate in thinking and non-thinking modes, with 1 million free tokens for 90 days. Beijing and Hong Kong endpoints list $1.413 / $4.24. Implicit cache hits cost $0.075 per million input tokens. Also available through Alibaba's Token Plan subscription. Weights are not downloadable. Check official pricing →

Quick Answer

Qwen4-Max is the top model in Alibaba’s Qwen family, released on November 16, 2026. It reads text, images, and video, writes text, has a 1M-token context window, and is built for coding, office work, front-end design, and long-running agent tasks. Its main selling point is price: $1.50 per million input tokens and $4.50 per million output on Alibaba’s international endpoint, well below most Western frontier models.

The catch most people will miss: unlike its predecessor Qwen3.8-Max, Qwen4-Max is not an open-weight model. You can use it through Alibaba’s API and apps, but you cannot download it and run it yourself.

Specifications

Qwen4-Max
DeveloperAlibaba (Qwen team), served by Alibaba Cloud Model Studio
ReleasedNovember 16, 2026 (preview from November 11)
API model IDqwen4-max (preview: qwen4-max-preview, retiring November 25, 2026)
ArchitectureSparse Mixture-of-Experts, 2.4 trillion total parameters, 71 billion active per token
Context window1,000,000 tokens; up to 991,808 input
Max output131,072 tokens (non-thinking); 262,144 including reasoning (thinking)
Input / outputText, images, and video in; text out
ReasoningThinking and non-thinking modes; effort low, high (default), xhigh
ToolsWeb search, tool calling, structured output (JSON object and JSON Schema)
API compatibilityOpenAI Chat Completions and Responses, Anthropic Messages, Alibaba DashScope
CachingImplicit and explicit caching (5-minute or 1-hour TTL)
WeightsNot released

Specifications are from Alibaba Cloud’s Model Studio documentation (last updated November 18, 2026) and Alibaba’s launch announcement.

Coding, Reasoning, and Agents

Alibaba aims Qwen4-Max at the same jobs as other frontier models: multi-step coding, front-end builds, professional knowledge work, and agents that run for hours. Its launch claims are worth reading in two groups.

Alibaba’s own figures. Alibaba says Qwen4-Max ranked sixth on Code Arena, a public coding leaderboard, and posted strong scores on Agents’ Last Exam, the Vals Index, JobBench, and Chartography. On Work360, a proprietary benchmark of industry tasks across 721 fields, it says Qwen4-Max beats Qwen3.8-Max while using 35% fewer tokens at high effort and cutting the average cost per task by about 60%. Alibaba also describes a 48-hour continuous agent run that designed and tested a coolant molecule. These are vendor results, and Work360 cannot be checked from outside.

Independent figures. Artificial Analysis scores Qwen4-Max 43 on its Intelligence Index, which it calls below average for models in a similar price tier (median 44), and well behind the leaders: Claude Fable 5.2 at 65, GPT-6.2 Astra at 62, and Gemini 4 Neon at 61. Kimi K3.1 Pro, the strongest open-weight model, scores 46. So the fair reading is a capable, inexpensive model rather than a contender for the top spot.

For agents, the practical features are there: tool calling, structured output, web search, and compatibility with both OpenAI’s Responses API and Anthropic’s Messages API, so it can slot into many agent frameworks and coding tools without custom adapters. In multi-turn use, thinking mode returns an encrypted reasoning record that you must pass back unchanged, as with other reasoning models.

Pricing and Availability

API. Through Alibaba Cloud Model Studio, Qwen4-Max is billed per token at a single rate for both thinking and non-thinking modes, with no long-context surcharge listed up to 1M tokens. Rates differ by deployment region:

EndpointInput (per 1M)Output (per 1M)Notes
Singapore (international)$1.50$4.501M free tokens, valid 90 days
Beijing, Hong Kong$1.413$4.24Batch in Beijing: $0.707 / $2.12
Cached input (implicit hit)$0.075Per Alibaba’s launch announcement

Model Studio also lists deployments in Frankfurt, Virginia, and Tokyo; check the pricing for your region and account in the console, since Alibaba’s tables vary by deployment scope. Third-party routers such as OpenRouter also list the model.

Subscriptions and apps. Qwen4-Max is included in Alibaba’s Token Plan, a subscription for using Qwen models in coding tools, and Alibaba says it is available in its QwenWork and QwenStudio apps.

The preview. qwen4-max-preview opened on November 11, 2026, through the Token Plan only, and Alibaba retires it after 24:00 on November 25, 2026 (UTC+8). Bills show it as qwen4.0-max-preview. Anyone who wired the preview into a tool should switch to qwen4-max now.

Licensing and Deployment Options

Qwen4-Max is a proprietary, API-accessible model. Use is governed by Alibaba Cloud’s service terms for the region you deploy in, not by an open model license. As of November 20, 2026, there are no Qwen4 weights in the Qwen organization on Hugging Face, and Alibaba’s announcement says only that “open-weight, smaller-scale models from the Qwen4 series” will be released later.

That has real consequences:

  • No self-hosting. You cannot run Qwen4-Max on your own servers or fine-tune the weights. If you need that, Kimi K3.1 Pro or Qwen3.8-Max are the large Chinese models with downloadable weights.
  • Data location depends on the endpoint. Choosing Singapore, Frankfurt, or Virginia changes where requests are processed. Read Alibaba Cloud’s data terms for that region before sending customer data, and check whether your organization allows a China-headquartered provider at all.
  • Vendor dependence. Prices, rate limits, and model retirement are Alibaba’s call, as the short-lived preview shows.

Practical Use Cases

  • High-volume coding help on a budget. At $4.50 per million output tokens, Qwen4-Max is cheap enough for routine code generation, test writing, and refactoring suggestions, with a stronger model as the escalation path for hard bugs. See model routing.
  • Long-document work. A 1M-token window with no listed long-context surcharge suits large contract sets, codebases, or research collections, if you verify what it cites.
  • Multimodal analysis. Reading charts, screenshots, and short videos alongside text, for example summarizing a recorded product demo with its slides.
  • Drop-in testing in existing tools. Because it speaks the OpenAI and Anthropic API formats, you can point an existing agent or coding tool at it and compare results against your current model on 20 real tasks.

Strengths

  • Low per-token price with a single rate across thinking and non-thinking modes
  • 1M-token context and large output limits
  • OpenAI- and Anthropic-compatible APIs, plus explicit and implicit caching
  • Text, image, and video input in one model
  • Several regional endpoints, including outside China

Limitations

  • Not open-weight, unlike Qwen3.8-Max and Kimi K3.1 Pro
  • Middling independent score. 43 on Artificial Analysis’s index, below the price-tier median
  • Many headline results are Alibaba’s own, including the proprietary Work360 benchmark
  • Regional differences in pricing, availability, and data handling that you have to sort out per account
  • Fast churn. The preview lasted two weeks; expect model IDs and terms to keep moving

Continue learning

Explore related guides, tools, workflows, and prompts that help you go deeper into this topic.

See how this tool fits into a workflow

Browse step-by-step AI workflows that use ChatGPT, Claude, Gemini, and other tools.

Frequently Asked Questions

Is Qwen4-Max open source or open-weight?

Neither, as of November 20, 2026. Alibaba offers Qwen4-Max only as a hosted model through its API, Token Plan, and apps, and no Qwen4 weights appear in the Qwen organization on Hugging Face. Alibaba says smaller open-weight Qwen4 models will follow. This differs from Qwen3.8-Max, whose text-only weights were published in August 2026.

How much does Qwen4-Max cost?

On Alibaba Cloud's international endpoint, $1.50 per million input tokens and $4.50 per million output tokens, in both thinking and non-thinking modes, with cached input at $0.075 per million on implicit cache hits. Endpoints in Beijing and Hong Kong list slightly lower rates ($1.413 / $4.24). New users of the international endpoint get 1 million free tokens valid for 90 days.

What is the difference between qwen4-max and qwen4-max-preview?

qwen4-max-preview was a limited preview that opened on November 11, 2026, available only through Alibaba's Token Plan with no pay-as-you-go option. Alibaba retires it after 24:00 on November 25, 2026 (UTC+8) and recommends the production qwen4-max model instead. Bills show the preview under its alias, qwen4.0-max-preview. If your code or agent config still names the preview, switch it before the deadline.

How big is Qwen4-Max's context window?

1,000,000 tokens, with up to 991,808 input tokens. Output is capped at 131,072 tokens in non-thinking mode and 262,144 tokens (reasoning plus answer) in thinking mode. Alibaba notes that thinking and the final answer share the same budget, so a low max_tokens setting can cut off the answer.

Is Qwen4-Max better than Kimi K3.1 Pro?

On Artificial Analysis's independent Intelligence Index, Kimi K3.1 Pro scores 46 and Qwen4-Max 43, so Kimi is slightly ahead on that measure. Qwen4-Max is cheaper per token ($1.50 / $4.50 against $2 / $10), while Kimi's weights can be downloaded and self-hosted. See our Qwen4-Max vs Kimi K3.1 Pro comparison.

Last updated: