Last updated: November 16, 2026
Qwen4-Max is Alibaba’s new flagship Qwen model, released on November 16, 2026. It is a 2.4 trillion parameter mixture-of-experts model with native vision and a 1 million token context window. You can call it as qwen4-max through the QwenCloud API at $1.50 per million input tokens, or use it through Token Plan.
Contents
- What Qwen4-Max is
- What changed from the preview
- Pricing and rate limits
- How to make your first call
- Thinking settings and output limits
- Limitations
- Frequently asked questions
What Qwen4-Max is
Alibaba’s launch announcement calls Qwen4-Max the most capable model in the Qwen family to date. It is a sparse mixture-of-experts model with 2.4 trillion total parameters, of which 71 billion are active per token. Alibaba says that is 25% fewer active parameters than Qwen3.8-Max at the same 2.4 trillion scale.
The Model Studio model page describes it as a multimodal model for coding, office work, frontend design and complex agent tasks. It accepts text, images and video, and it returns text. Supported features include:
- Deep thinking, with a thinking summary and encrypted reasoning for multi-turn use.
- Function calling, parallel tool calling and built-in web search.
- Structured output with JSON Object and JSON Schema.
- Explicit and implicit context caching, streaming and prefix completion.
- OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and DashScope protocols.
Alibaba also reports its own results. It says Qwen4-Max ranked 6th on Code Arena. On Work360, its internal benchmark covering 721 specialized fields, Alibaba says the model beat Qwen3.8-Max overall while using 35% fewer tokens at High effort and cutting average cost per task by around 60%. These are vendor claims, not independent tests.
Besides the API, Alibaba lists access in QwenWork and QwenStudio. For the desktop app, see our QwenWork 2.0 guide.
What changed from the preview
The model first appeared as Qwen4-Max-Preview. The production model differs in several ways that matter if you built against the preview:
| Item | qwen4-max-preview (Nov 11) | qwen4-max (Nov 16) |
|---|---|---|
| Access | Token Plan only; no pay-as-you-go | Pay-as-you-go API and Token Plan |
| Status | Retired after November 25, 2026, 24:00 (UTC+8) | Production model, recommended by Alibaba |
| Output budget | 393,216 tokens shared by thinking and answer | 262,144 tokens for thinking plus answer; 131,072 without thinking |
| Effort values | low, medium, xhigh (default xhigh) | low, high, xhigh (default high) |
| Published price | Token Plan Credits only | $1.50 input / $4.50 output per 1M tokens (Singapore) |
Sources: the preview page, the Qwen4-Max page and the QwenCloud model releases log. The Token Plan updates page confirms that both Personal Edition and Team Edition support Qwen4-Max as of November 16.
Pricing and rate limits
The model pricing page lists one flat tier for requests up to 1 million input tokens, covering both thinking and non-thinking modes:
| Region | Input per 1M tokens | Output per 1M tokens | Free quota |
|---|---|---|---|
| Singapore (International) | $1.50 | $4.50 | 1 million tokens, valid 90 days |
| China (Beijing), Hong Kong, Frankfurt, Virginia, Tokyo | $1.413 | $4.24 | None listed |
Cache pricing in Singapore is $1.875 per million tokens for a 5-minute cache write, $3 for a 1-hour cache write and $0.075 for explicit or implicit cache reads. For comparison, the same page lists qwen3.8-max at $2 input and $6 output in Singapore, so the new flagship is cheaper per token than its predecessor.
Rate limits depend on region. In Singapore and Beijing, the tokens-per-minute limit is 5 million at Tier 1 (monthly spend up to $8,000), 10 million at Tier 2 and 20 million at Tier 3 (over $140,000). The Global regions list 30,000 requests and 5 million tokens per minute.
How to make your first call
The Model Studio quickstart uses the standard OpenAI and Anthropic SDKs with a workspace-specific endpoint:
- Get a Model Studio API key and store it in the
DASHSCOPE_API_KEYenvironment variable, which the quickstart examples read. - Set the base URL to
https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1for OpenAI-style calls, replacing the placeholder with your workspace ID. - For the Anthropic SDK, use
https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/apps/anthropicinstead. - Set
modeltoqwen4-maxand send a request through Responses, Chat Completions or Messages. - If you migrate from the preview, change the model ID and recheck any output limits sized for the larger preview budget.
If you prefer a subscription for coding tools, the Token Plan overview shows a one-line setup with the Model Studio CLI, for example bl config agent --agent claude-code --region ap-southeast-1 --api-key $TOKEN_PLAN_API_KEY --model auto. The page shows the same pattern for codex, qwen-code, openclaw and hermes. For Alibaba’s own terminal agent, see our Qwen Code 1.0 Preview guide.
Thinking settings and output limits
Thinking is on by default. The documented defaults are enable_thinking=true, preserve_thinking=true, reasoning_effort=high and thinking_budget=131072. Effort accepts low, high and xhigh. Other values are mapped: minimal becomes low, medium becomes high and max becomes xhigh. In the Anthropic Messages API, the setting is output_config.effort.
Context limits from the model page:
- Context length: 1,000,000 tokens.
- Maximum input: 983,616 tokens with thinking, 991,808 without.
- Thinking plus final answer: up to 262,144 tokens.
- Non-thinking output: up to 131,072 tokens.
Default sampling also changes by mode: temperature 1.0 and top_p 0.95 with thinking, 0.7 and 0.8 without. The preview notes add that, starting with qwen4-max-preview, max_tokens in Chat Completions limits thinking and the answer together, so leave room for both. In multi-turn chats, pass encrypted reasoning back unchanged.
Limitations
- No open weights. Alibaba says smaller open-weight Qwen4 models will come later, with no date.
- No batch or fine-tuning. The model page lists batch inference, fine-tuning and model deployment as not supported.
- Text output only. It reads images and video but does not generate them.
- Vendor benchmarks. The Code Arena and Work360 claims come from Alibaba.
For other Chinese flagship models, compare our Kimi K3.1 Pro guide and LongCat-2.5 guide. For free consumer access, see Is Qwen free?
Frequently asked questions
When was Qwen4-Max released?
Alibaba released Qwen4-Max on November 16, 2026. It follows Qwen4-Max-Preview, which opened on November 11, 2026 on Token Plan only and will be retired after its preview period ends on November 25, 2026 (UTC+8).
How much does the Qwen4-Max API cost?
In the Singapore (International) region, Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens. Other listed regions charge $1.413 and $4.24. Cache reads cost $0.075 per million tokens, and Singapore includes a 1 million token free quota.
What is the Qwen4-Max context window?
Qwen4-Max has a 1,000,000-token context window. Thinking and the final answer together can use up to 262,144 tokens, and non-thinking output is capped at 131,072 tokens, according to Alibaba Cloud Model Studio documentation.
Is Qwen4-Max open source?
No open weights have been released for Qwen4-Max. Alibaba says smaller open-weight models from the Qwen4 series will be released to developers worldwide soon, but it has not given a date.
Can I use Qwen4-Max with Claude Code or Codex?
Yes. Qwen4-Max supports the Anthropic Messages and OpenAI APIs, and Token Plan Personal and Team Editions support the model. Alibaba's Model Studio CLI can configure Claude Code, Codex, Qwen Code and other agents with a Token Plan API key.
Sources
- qwen4-max (Latest Model), Alibaba Cloud Model Studio documentation
- Alibaba Unveils Qwen4-Max, Alibaba Cloud Community (November 16, 2026)
- Model pricing, Alibaba Cloud Model Studio documentation
- qwen4-max-preview, Alibaba Cloud Model Studio documentation
- Token Plan product updates, Alibaba Cloud Model Studio documentation