Qwen4-Max: API Access, Pricing, Free Quota, and What to Watch

Qwen4-Max is Alibaba’s newly released multimodal reasoning model. It is available now through QwenCloud’s pay-as-you-go API and Token Plan, as well as QwenStudio and QwenWork. The API has published prices; eligible users can receive a limited free quota for real-time calls in Singapore. Use the production model ID qwen4-max.

Can You Use Qwen4-Max Now? How?

Yes. Create a QwenCloud API key for your chosen region, then call qwen4-max through its OpenAI-compatible Chat Completions API. The official quick start shows the Singapore base URL https://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1. It also documents Responses, Anthropic Messages, and DashScope access. Token Plan calls use separate credentials and an endpoint specified in its plan guide.

The model accepts text, images, and video as input and returns text. Its context window is 1,000,000 tokens. Maximum output is 131,072 tokens without thinking, or 262,144 tokens for thinking and final answer combined. The maximum input is lower than the full context window and varies by thinking mode. These figures come from the official model specification.

Qwen4-Max API Pricing and Free Quota

Pay-as-you-go prices vary by deployment region. The figures below are US dollars per million tokens from QwenCloud’s pricing table. Thinking tokens are billed as output tokens.

Billing itemSingapore (International)Beijing / Global
Input$1.50$1.413
Output, including thinking$4.50$4.24
Cached input read$0.075$0.071

The model guide lists a 1,000,000-token free quota for eligible Singapore pay-as-you-go calls. The free-quota rules say it covers real-time inference through a Singapore API key and endpoint, not built-in tool fees or Token Plan calls. Payment information is required to activate the quota. Once it runs out, pay-as-you-go billing can begin. Check the model’s balance and expiration in the Free Tier console, and enable Free quota only there if you want calls to stop when the quota ends.

What Changes From Qwen3.8-Max?

The clearest directly comparable change is the Singapore list price: the Qwen3.8-Max model page lists $2 input and $6 output per million tokens, while the Qwen4-Max pricing table lists $1.50 and $4.50. Both official model pages list a 1,000,000-token context window, so context size alone is not the upgrade. Alibaba says its own Work360 evaluation found lower token use at high reasoning effort; that is a vendor evaluation, not a result established for every workload.

Thinking Is On by Default

QwenCloud says thinking is enabled by default with reasoning_effort=high and a default thinking_budget of 131,072 tokens. In Chat Completions, set enable_thinking=false to turn thinking off. Since thinking tokens count as output tokens, this setting can affect the bill as well as response behavior.

Cache Prices and a Discount Limit

In Singapore, the published price per million tokens is $0.075 for a cache read, $1.875 for a five-minute explicit cache write, and $3 for a one-hour explicit cache write. Beijing and Global prices differ; consult the regional price table before calling those endpoints. QwenCloud also says batch and cache discounts cannot be combined on the same request. Its Qwen4-Max specification currently marks batch inference unsupported, so do not assume a general batch discount applies to this model.

Failed Calls Do Not Consume the Free Quota

QwenCloud’s billing documentation states that failed API calls are not charged and do not consume free quota. That rule concerns failed calls; a successful call still uses its applicable tokens or credits.

Current Catches and an Early User Test

  • API choice matters for caching. The model guide says Qwen4 explicit caching works through Chat Completions, Responses, and Anthropic Messages, but not DashScope. DashScope uses its multimodal endpoint even for text-only Qwen4-Max calls.
  • Free quota has a scope. The official rules exclude built-in tool fees and calls outside Singapore. The Free quota only switch is disabled by default.
  • User test, not an official guarantee. In a post-launch Hacker News comment, one user wrote: “If you use the same prompt via the qwenapp or qwen studio you get an interactive web page with an animated scene.” This describes one run in those interfaces, not a verified API result.

Frequently Asked Questions

Do I need a Token Plan?

No. The production qwen4-max model also supports pay-as-you-go API access. The separate qwen4-max-preview has different access rules. See the model guide and release changelog.

Can I download Qwen4-Max weights?

Alibaba’s announcement describes hosted Qwen4-Max access and promises open weights for smaller Qwen4 models. It does not announce downloadable Qwen4-Max weights or a release date for those smaller models.

What are the API rate limits?

Singapore and Beijing use dynamic tiers. The Global regions list 30,000 requests per minute and 5,000,000 tokens per minute. Check the official rate-limit page and your account’s current limits before deployment.

Can I fine-tune or deploy the model myself?

No. The Qwen4-Max specification marks fine-tuning and model deployment unsupported.