On November 11, Alibaba's Qwen team announced on its official account that Qwen4-Max is entering preview: the first model of the Qwen 4 family anyone can call, live for just 15 days. The preview runs from 10:00 on November 11 to 24:00 on November 25, 2026 (UTC+8), after which the model is retired, with no changes in between.
According to Alibaba Cloud Model Studio documentation, Qwen4-Max-Preview is the next-generation flagship multimodal model of the Qwen family, supporting text generation, image and video understanding, deep reasoning and tool calling, with a 1-million-token context and a 393,216-token ceiling on a single output, shared between the reasoning trace and the final answer. Its API is compatible with OpenAI and Anthropic formats. There are five places to try it: the Token Plan subscription, QwenWork, Qoder, the Qwen App and QwenStudio — and only the Token Plan is a developer API route. Pay-as-you-go is not offered, and the service runs in the Singapore region only.
No benchmarks, no price — hold off on conclusions
What stands out about this preview is not what Alibaba published, but what it did not: no parameter count, no weights, no license, no per-token price and no benchmark score at all. The Qwen team lists four capability improvements with no benchmark numbers behind them. The Token Plan Individual tier starts at $3 a month (Lite event price, list $8), but that is a subscription price, not a per-call price for the model. In other words, the only thing testable right now is whether it is pleasant to use; how strong it is cannot be judged yet. The Qwen family has been busy lately: Qwen-Image-2.1-Turbo generates images in 8 steps focuses on image generation, while Qwen4-Max covers the understanding and reasoning side.
The Qwen 4 family was first shown on September 22 at the Apsara conference with four tiers named and no release dates; on October 29 Alibaba Cloud's CTO said it was a week away, and on November 11 the Max tier arrived as a preview. When the other tiers follow, when a stable ID replaces the preview ID, and when benchmarks appear are all still unknown. Once the 15-day window closes, the qwen4-max-preview ID disappears — anyone who wants to try it has half a month.
