Which region you call it from changes the price of qwen4-max. Before you read anything else about Alibaba's new flagship, decide where your requests will originate, because the answer moves the bill by a few percent and decides whether batch discounts exist at all.
General availability arrived on Alibaba Cloud Model Studio on 16 November 2026, after a short Token Plan preview. Some coverage dates it 15 November, which is the same moment in US time zones. It is the first model of the Qwen4 generation, and the weights are not released; Alibaba says smaller Qwen4 models will be open-sourced later, without naming them.
Price by region
Indicative USD per 1M tokens, one flat tier up to the full 1M input:
| Region |
Input |
Output |
Cache read |
Cache write (5 min / 1 h) |
| Singapore (international) |
$1.50 |
$4.50 |
$0.075 |
$1.875 / $3.00 |
| Beijing, Hong Kong, Frankfurt, Virginia, Tokyo |
$1.413 |
$4.24 |
$0.071 |
$1.767 / $2.827 |
Batch inference runs only in Beijing, at $0.707 input and $2.12 output. Singapore accounts get a free quota of 1M tokens valid for 90 days. Confirm all of these on your own account; Model Studio prices change.
Rate limits differ by region too. Singapore and Beijing are tiered by monthly spend, starting at 5M tokens per minute. Frankfurt, Virginia, Hong Kong and Tokyo list 30,000 requests and 5M tokens per minute.
Token budgets, precisely
The context window is 1,000,000 tokens, but the usable split depends on mode.
- Thinking mode (default): up to 983,616 input tokens and 262,144 output tokens.
thinking_budget defaults to 131,072.
- Non-thinking: up to 991,808 input tokens and 131,072 output tokens.
Listings that give 262,144 as the maximum output are quoting the thinking-mode figure. If you switch thinking off, the ceiling halves.
reasoning_effort accepts low, high (default) and xhigh. Aliases map minimal to low, medium to high and max to xhigh. The Responses API uses reasoning.effort; the Anthropic Messages interface uses output_config.effort.
Interfaces and features
You can call it through OpenAI Chat Completions, OpenAI Responses, Anthropic Messages or DashScope. Input is text, image or video; output is text.
Supported: streaming, parallel tool calling, function calling, web search, structured output with JSON Schema, prefix completion, explicit and implicit caching, and thinking summaries. Fine-tuning and dedicated deployment are not available.
Default sampling differs by mode. Thinking uses temperature 1.0 and top_p 0.95. Non-thinking uses 0.7 and 0.8, with a presence penalty of 1.5 that can surprise you on repetitive structured output.
On the parameter count
Several blogs describe a 2.4T-parameter backbone with 71B active. No Alibaba page we read states a parameter count, so treat that figure as unconfirmed.
Trying it against what you run now
Start with thinking on at low effort. Many coding and document tasks will not need more, and the default budget is generous enough to inflate costs quietly. Then raise effort only on the prompts that fail. Compare against Qwen3.8 Max on the same set, since that is the model most teams will be upgrading from. If you held a Token Plan and used the preview, migration is short: switch the model name and move to per-token billing. The rest of the family is on the Alibaba provider page.
For broader context, browse reasoning-capable text models.