Alibaba Opens a Preview of Qwen4-Max With a Million Token Window
Alibaba has opened a preview of Qwen4-Max, the next flagship in its Qwen model family, and the only way to reach it is a subscription plan rather than a per token bill.
The preview runs on Alibaba's Token Plan until the end of this month, carries a context window of one million tokens, and accepts images and video as input alongside text. The company says the model will not change during the window and will be retired when it closes.
And the specification is unusually complete for a preview. Output is capped at 393,216 tokens shared between reasoning and the final answer, thinking is on by default and can be adjusted through a reasoning effort setting, and the model supports web search, function calling, built in tools and structured JSON.
The interfaces are deliberately familiar. Model Studio documentation lists compatibility with OpenAI's chat completions and responses formats and with Anthropic's messages API, which means existing code can be repointed at the Qwen4 endpoint with little more than a model identifier change.
Access is tight by design. The preview is not available on pay as you go, calls draw on Token Plan credits at a promotional rate the company lists as low as 20 per cent of standard, and the same model is switched on inside QwenWork, Qoder, the Qwen app and QwenStudio.
But the gap between the documentation and the claims is wide. Alibaba has published no benchmark results, no model card, no parameter count, no price per token and no weights, and has not dated the full Qwen4 launch, so the assertions about deeper knowledge, richer perception and stronger coding are the company's own.
The marketing language also hides a real question for buyers. A one million token window and a fixed retirement date make the preview useful for testing inside coding and agent tools, but unsuitable for anything that has to keep working after the window closes.
Alibaba said the Qwen4 family entered training earlier this year, and the naming suggests a deliberate split between this preview and whatever ships as the finished model, since the company has previously released incremental versions of its flagships before a full generation.
So the preview is best read as a controlled beta for developers who already pay for Token Plan, and the number that matters is not the context window but what Alibaba publishes when it has benchmark figures to defend.
Alibaba opened a preview of Qwen4-Max on its Token Plan
The context window is one million tokens with output capped at 393,216
It accepts image and video input and supports tool calling and JSON output
The model is not available on pay as you go and retires at the end of this month
No benchmarks, parameter count, pricing or weights have been published
Source: Pandaily, Alibaba Cloud Model Studio
