Alibaba Starts Two-Week Qwen4-Max Preview and Withholds Benchmarks
Alibaba spent the summer opening the weights of its last flagship model and previewing the architecture of the next one. It is now putting Qwen4-Max in front of users without giving them either.
The official Qwen account said on November 11 that Qwen4-Max is in preview on Alibaba’s Token Plan, with early access also available on QwenWork, Qoder, Qwen APP, and QwenStudio. The post offered no parameter count, no official benchmarks, and no generally published API price. Alibaba Cloud’s Token Plan documentation, updated the same day, is more specific about the terms: the preview runs until midnight on November 25, 2026, Beijing time, and the preview model will then be retired.
That is a tighter release than the one Alibaba used for Qwen3.8-Max in early August. The company published the model’s scale — 2.4 trillion parameters, with 95 billion active — and, at launch, attached public scores, saying it ranked fifth on Text Arena and second on Vision Arena. The model was offered through Alibaba Cloud’s Model Studio API, and the weights followed on Hugging Face and ModelScope. Later that month, Qwen open-sourced Qwen3.8-Flash-Next as an early look at the architecture intended for Qwen4, a move ChinaBiz Insider covered at the time. The company was willing, in other words, to put a flagship model’s size and performance in public, and to let the community inspect the next architecture before the new family arrived.
Qwen4-Max is the first time a Max-branded model in that new generation has been placed in users’ hands. In this preview, Alibaba has put early access inside its own products and paid subscription, rather than first handing the model to the open-source community. Developers can try it. They cannot download it, and they cannot compare it on a public leaderboard. After two weeks, according to the official docs, the preview comes down.
The commercial gate is Token Plan, a credit-based subscription for individual developers, currently offered only in the Singapore region and meant for interactive use inside coding and agent tools. It is not a traditional pay-as-you-go API. During the preview, credit consumption can be as low as 20% of the standard rate, according to Alibaba Cloud’s documentation. Those figures measure burn inside the subscription, not a published API list price, and the company says it may change or end the promotion.
The documentation identifies the model as qwen4-max-preview, also callable as qwen4.0-max-preview, and lists its capabilities as reasoning, visual understanding, and text generation. The official launch post said the new generation would offer “deeper knowledge,” “richer world understanding,” and “stronger coding and agentic capabilities.” That is a product claim, not a measurement. Until official numbers appear, comparisons with Anthropic, OpenAI, or Google models will be anecdotal.
For Alibaba, Qwen4-Max marks a different kind of launch. The company is opening access to the model, but keeping the model itself behind its own products and subscription system. Whether that approach will extend to the full Qwen4 release remains to be seen.
Related Coverage：
Alibaba’s Qwen3.8-Flash Rewrites the AI Cost Curve, Previews Qwen4 Architecture