# Qwen4-Max: price, API access and 1M context

> Alibaba's Qwen4-Max is a 2.4T MoE flagship with 71B active parameters, native vision and a 1M-token context. See the API ID, limits and regional prices.

- Canonical HTML: https://omniakey.com/blog/qwen4-max
- Markdown: https://omniakey.com/blog/qwen4-max.md
- Published: 2026-11-16 (updated 2026-11-16)
- Author: Kiki, COO at OmniaKey
- Language: en
- Category: Guide
- Reading time: 8 min
- All posts: https://omniakey.com/blog/llms.txt

---

Qwen4-Max is available now. Alibaba released it on November 16, 2026, and QwenCloud sells the model as `qwen4-max` through the Token Plan and the pay-as-you-go API, with Singapore input at $1.50 per million tokens. The release lists 2.4 trillion total parameters with 71 billion active per token, native vision and a 1M-token context ([Alibaba, November 16, 2026](https://www.alibabacloud.com/en/press-room/alibaba-unveils-qwen4-max?_p_lc=1); [QwenCloud latest model](https://docs.qwencloud.com/developer-guides/getting-started/latest-model); [QwenCloud pricing](https://docs.qwencloud.com/developer-guides/getting-started/pricing)).

**Short answer**

- Qwen4-Max is live as `qwen4-max` on QwenCloud, through Token Plan Personal, Token Plan Team or the pay-as-you-go API ([QwenCloud](https://docs.qwencloud.com/developer-guides/getting-started/latest-model)).
- Singapore pay-as-you-go lists $1.50 per million input tokens and $4.50 per million output tokens; Beijing and the Global regions list $1.413 and $4.24 ([pricing](https://docs.qwencloud.com/developer-guides/getting-started/pricing)).
- The context window is 1M tokens, and thinking plus the final answer can reach 262,144 tokens.
- Qwen4-Max has no open-weight download. Alibaba says smaller Qwen4 open-weight models will follow "soon" ([release](https://www.alibabacloud.com/en/press-room/alibaba-unveils-qwen4-max?_p_lc=1)).

## Is Qwen4-Max available now?

Yes. Alibaba announced the model in Hangzhou on November 16, 2026, and says it is available on QwenCloud through the Token Plan and API, plus Alibaba's own QwenWork and QwenStudio apps ([release](https://www.alibabacloud.com/en/press-room/alibaba-unveils-qwen4-max?_p_lc=1)). QwenCloud's docs name three ways to pay: Token Plan Personal, Token Plan Team, or pay-as-you-go ([QwenCloud](https://docs.qwencloud.com/developer-guides/getting-started/latest-model)). Alibaba Cloud's Model Studio index lists `qwen4-max` under complex reasoning and coding and under visual understanding, on a page last updated November 16, 2026 ([Model Studio](https://www.alibabacloud.com/help/en/model-studio/models)).

Deployment covers Singapore (International), China North 2 in Beijing, and Global in Frankfurt, Hong Kong, Tokyo and US Virginia. Singapore uses the base URL `https://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1` ([QwenCloud](https://docs.qwencloud.com/developer-guides/getting-started/latest-model)). Pay-as-you-go calls in Singapore include 1 million free tokens ([QwenCloud pricing](https://docs.qwencloud.com/developer-guides/getting-started/pricing)), separate from Token Plan credits.

Coding tools get an explicit listing: QwenCloud names Claude Code, Codex and OpenCode ([QwenCloud](https://docs.qwencloud.com/developer-guides/getting-started/latest-model)). For the custom-endpoint steps those tools need, see the [Claude Code setup guide](https://omniakey.com/blog/claude-code-setup), and the [coding agent comparison](https://omniakey.com/blog/best-llm-coding-agents) for the wider field.

Status checked November 16, 2026.

## What changed from Qwen3.8-Max?

Qwen4-Max is a sparse mixture-of-experts (MoE) model. It keeps the 2.4 trillion total parameter scale of Qwen3.8-Max and cuts activated parameters by 25%, to 71 billion per token ([release](https://www.alibabacloud.com/en/press-room/alibaba-unveils-qwen4-max?_p_lc=1)). Alibaba describes the MoE change as covering attention, residual connections and embeddings. QwenCloud calls the model the first in the Qwen4 family and puts reasoning, native vision and long-running task execution in one sentence ([QwenCloud](https://docs.qwencloud.com/developer-guides/getting-started/latest-model)).

Both models sit at the same 2.4T total scale. The total and active counts describe compute per token, while Work360 is Alibaba's separate vendor-reported task result.

Alibaba reports sixth place on Code Arena, a third-party coding leaderboard, and says Qwen4-Max used 35% fewer tokens at High effort than Qwen3.8-Max on Work360 while cutting average cost per task by around 60% ([release](https://www.alibabacloud.com/en/press-room/alibaba-unveils-qwen4-max?_p_lc=1)). Work360 is Alibaba's proprietary benchmark, scored across 721 specialized fields. The same release names Agents' Last Exam, Vals Index, JobBench and Chartography without publishing scores.

Those figures come from Alibaba, and no independent evaluation appeared in the sources checked. Read the Code Arena placement as Alibaba's statement about a third-party board, and the Work360 numbers as a vendor benchmark that measures Alibaba's own tasks.

## Qwen4-Max specs and model ID

The production ID is `qwen4-max` ([QwenCloud](https://docs.qwencloud.com/developer-guides/getting-started/latest-model)). `qwen4-max-preview` is a separate entry with its own ID, billing and retirement date ([Model Studio](https://help.aliyun.com/en/model-studio/qwen4-max-preview)). Input is text, image and video; output is text ([QwenCloud model table](https://docs.qwencloud.com/developer-guides/getting-started/text-generation-models)).

|Limit|Value|
|-|-|
|Context window|1,000,000 tokens|
|Max input, non-thinking|991,808 tokens|
|Max input, thinking|983,616 tokens|
|Max output, non-thinking|131,072 tokens|
|Thinking plus answer|262,144 tokens|
|Default thinking budget|131,072 tokens|

Thinking is on by default with `reasoning_effort=high` and a 131,072-token thinking budget ([QwenCloud](https://docs.qwencloud.com/developer-guides/getting-started/latest-model)). The effort setting takes `low`, `high` or `xhigh`, and QwenCloud maps `minimal`, `medium` and `max` onto those three values. In Chat Completions, `enable_thinking=false` turns thinking off. `max_completion_tokens` caps thinking and the final answer together, so a thinking request can spread up to 262,144 tokens across both.

Batch inference is not supported for `qwen4-max` ([QwenCloud](https://docs.qwencloud.com/developer-guides/getting-started/latest-model); [model table](https://docs.qwencloud.com/developer-guides/getting-started/text-generation-models)), and neither is fine-tuning or model deployment. The APIs are Chat Completions, Responses, Anthropic Messages and DashScope, and text-only requests still use the DashScope multimodal endpoint.

Media limits allow 16 million pixels per image, up to 2,048 image URLs, 250 base64 images and 64 videos per request, with video up to 2 hours or 2GB ([QwenCloud model table](https://docs.qwencloud.com/developer-guides/getting-started/text-generation-models)). Those caps describe what the endpoint accepts, while the 991,808-token input ceiling decides how much of it fits in one call.

## Qwen4-Max API pricing by region

QwenCloud's pricing page, checked November 16, 2026, lists these pay-as-you-go rates in US dollars per million tokens. Thinking and non-thinking modes cost the same, and Qwen4-Max has no input-length tiers ([QwenCloud pricing](https://docs.qwencloud.com/developer-guides/getting-started/pricing)).

|Billing item|Singapore|Beijing / Global|
|-|-|-|
|Input|$1.50|$1.413|
|Output, thinking + answer|$4.50|$4.24|
|Cache write, 5m|$1.875|$1.767|
|Cache write, 1h|$3|$2.827|
|Cache read|$0.075|$0.071|

Two details change the arithmetic. Thinking tokens are billed as output tokens ([QwenCloud pricing](https://docs.qwencloud.com/developer-guides/getting-started/pricing)), and cache writes and reads use the rates above. Explicit caching takes a 5-minute or 1-hour time-to-live, with 5 minutes as the default; the new Qwen4 explicit cache is not available on DashScope ([QwenCloud](https://docs.qwencloud.com/developer-guides/getting-started/latest-model)).

A worked example for Singapore: 100,000 input tokens at $1.50 per million is $0.15, and 20,000 output tokens at $4.50 per million is $0.09, so one request lands near $0.24. The same request in Beijing or Global comes to about $0.23.

Batch pricing at 50% of real-time rates does not apply here, because batch inference is unsupported for this model ([pricing](https://docs.qwencloud.com/developer-guides/getting-started/pricing)). Built-in tools bill on top: web search is $10 per 1,000 calls in Singapore and $0.573411 in Beijing and Global, while function calling and MCP carry no tool fee. Token Plan credit rates and per-model rate limits are not published on the pages checked.

## What can Qwen4-Max do with vision and tools?

Image and video understanding both appear as supported capabilities, and the same entry gets the 1M-token context ([QwenCloud](https://docs.qwencloud.com/developer-guides/getting-started/latest-model)). Input is text, image and video, with text output ([model table](https://docs.qwencloud.com/developer-guides/getting-started/text-generation-models)).

Tool support covers function calling, parallel tool calls, web search, web scraping, code interpreter, text-to-image search, image-to-image search, PDF understanding and MCP ([QwenCloud](https://docs.qwencloud.com/developer-guides/getting-started/latest-model)). The code interpreter needs thinking enabled. Structured output runs as JSON Object or JSON Schema, and streaming, prefix completion and implicit caching are supported. CoT Summary returns a readable summary while the full reasoning stays encrypted.

`preserve_thinking` is on by default for multi-turn tool calls. That setting matters for agent loops: earlier reasoning stays available to later turns unless you switch it off.

Alibaba's release adds customer and internal examples. A cloud scheduling test cut data-annotation model call costs by 9% and made offline CPU rescheduling more than 14 times faster. An automotive customer cut manual workpiece programming from about 475 minutes to 24. A science workflow ran 48 hours through four closed-loop iterations, and an Alzheimer's research run made 454 tool calls and 3,239 metric evaluations to produce 24 candidate molecules ([release](https://www.alibabacloud.com/en/press-room/alibaba-unveils-qwen4-max?_p_lc=1)). Those results are Alibaba's own reports.

## Is Qwen4-Max open source?

No. The release, the QwenCloud docs and the Model Studio pages publish no weights, download link or licence for Qwen4-Max. Alibaba says open-weight, smaller-scale models from the Qwen4 series "will soon be released to developers worldwide" ([release](https://www.alibabacloud.com/en/press-room/alibaba-unveils-qwen4-max?_p_lc=1)). Until those appear, `qwen4-max` is a hosted model reachable through QwenCloud, Alibaba's apps and Model Studio.

## Qwen4-Max or Qwen4-Max-Preview?

Use `qwen4-max` for new work. The preview entry `qwen4-max-preview` is Token Plan only, with no pay-as-you-go access ([Model Studio](https://help.aliyun.com/en/model-studio/qwen4-max-preview)). Its preview runs from 10:00 on November 11, 2026 to 24:00 on November 25, 2026 (UTC+8), and the model is retired afterward. The preview caps output at 393,216 tokens shared by thinking and answer, against 262,144 for production in thinking mode. Alibaba recommends the production version.

One numbering detail carries over: from `qwen4-max-preview` onward, `max_tokens` means the same as `max_completion_tokens`. Code written for older Qwen models can spend more than intended if it assumes the earlier meaning.

> Preview keys stop working after November 25, 2026 at 24:00 UTC+8. Move scheduled jobs and long-running agents to&#x20;
>
> `qwen4-max`
>
> &#x20;before that date.

The preview's Token Plan base URL is `https://token-plan.cn-beijing.maas.aliyuncs.com` ([Model Studio](https://help.aliyun.com/en/model-studio/qwen4-max-preview)); the production model uses the QwenCloud endpoint for its region. For the preview's access tiers and Qoder details, see the [Qwen4-Max-Preview access guide](https://omniakey.com/blog/qwen4-max-preview).

## How to choose an access path today

|Situation|Action|
|-|-|
|New integration|Singapore pay-as-you-go, 1M free|
|Token Plan team|Switch ID to `qwen4-max`|
|Preview key in a job|Move before November 25|
|Batch jobs|Keep the current model|
|One key, many models|OmniaKey path below|

OmniaKey's catalog carries [Qwen3.8-Max](https://omniakey.com/qwen3-8-max), the model Alibaba compares Qwen4-Max against, and does not list Qwen4-Max. For a Qwen model through one key today, Qwen3.8-Max is the option in the catalog; the [pricing page](https://omniakey.com/pricing) has current rates.

Check the model list in your own account before you commit a long job to either ID, and price a real request from your own token counts rather than from average token figures.

## Frequently asked questions

### Can I use Qwen4-Max in Claude Code?

Alibaba's QwenCloud docs list Claude Code, Codex and OpenCode under coding tools for `qwen4-max` ([QwenCloud](https://docs.qwencloud.com/developer-guides/getting-started/latest-model)). Use a QwenCloud key and the endpoint for your region; the general custom-endpoint steps are in the [Claude Code setup guide](https://omniakey.com/blog/claude-code-setup).

### Does Qwen4-Max support batch inference?

No. QwenCloud lists batch inference as not supported for `qwen4-max`, so the 50% batch rate on its pricing page does not apply ([model table](https://docs.qwencloud.com/developer-guides/getting-started/text-generation-models); [pricing](https://docs.qwencloud.com/developer-guides/getting-started/pricing)). Send large jobs as separate requests, or keep them on a model that accepts batch calls.

### What is the maximum output for Qwen4-Max?

131,072 tokens in non-thinking mode, and 262,144 tokens shared by thinking and answer when thinking is on ([QwenCloud](https://docs.qwencloud.com/developer-guides/getting-started/latest-model)). `max_completion_tokens` caps both parts together.

### What happens to qwen4-max-preview on November 25?

The preview retires at 24:00 UTC+8 on November 25, 2026, and Alibaba points users to the production `qwen4-max` model ([Model Studio](https://help.aliyun.com/en/model-studio/qwen4-max-preview)). Rename any job that still uses the preview ID.

### Where can I use Qwen4-Max through OmniaKey today?

OmniaKey does not list Qwen4-Max yet. The catalog's Qwen model is [Qwen3.8-Max](https://omniakey.com/qwen3-8-max), available with one OmniaKey key; current rates are on the [pricing page](https://omniakey.com/pricing).

## Sources checked

- [Alibaba: Alibaba unveils Qwen4-Max (November 16, 2026)](https://www.alibabacloud.com/en/press-room/alibaba-unveils-qwen4-max?_p_lc=1)
- [QwenCloud: Latest model](https://docs.qwencloud.com/developer-guides/getting-started/latest-model)
- [QwenCloud: Pricing (checked November 16, 2026)](https://docs.qwencloud.com/developer-guides/getting-started/pricing)
- [QwenCloud: Text generation models](https://docs.qwencloud.com/developer-guides/getting-started/text-generation-models)
- [Alibaba Cloud Model Studio: Models](https://www.alibabacloud.com/help/en/model-studio/models)
- [Alibaba Cloud Model Studio: qwen4-max-preview](https://help.aliyun.com/en/model-studio/qwen4-max-preview)
