Qwen4-Max Is Now in Preview on Alibaba's Token Plan

Alibaba has moved Qwen4 from a teaser to something developers can actually try. The official @Alibaba_Qwen post leads with "Qwen4 is coming," then says the flagship, Qwen4-Max, is now in preview on Alibaba's Token Plan, with early access also available on QwenWork, Qoder, the Qwen APP, and QwenStudio. The pitch is compact: deeper knowledge, richer world understanding, and stronger coding and agentic capabilities, aimed at complex coding, knowledge work, front-end builds, and professional software workflows.

This is the first time the company has officially described Qwen4-Max as in preview, rather than still on the way. The post does not include a model card, a parameter count, a context window, an API model ID, or official benchmarks. What is confirmed is where to try it, not what the model is made of.

A preview of the flagship is not a launch of the whole family

Qwen4 has been described as already in training since this year's Apsara Conference, and the lineup names Max, Plus, Flash, and 27B have been public for some time. The only model officially placed in preview today is the flagship, Qwen4-Max. The opening line, "Qwen4 is coming," reads more like a notice that the generation is still rolling out than a claim that Plus, Flash, or an open-weight 27B have arrived with it.

It is also not a general availability release, and it is not an open-weight drop. The official language is preview and early access. No license, no downloadable checkpoint, and no system card appear in today's announcement.

A stealth, unnamed model showed up earlier in Qoder, and the community widely guessed it was Qwen4. Today's post does name Qwen4-Max and does list Qoder as an early-access surface, but it never identifies that stealth name. Until Alibaba makes the mapping itself, the two should not be treated as the same object. Circulating figures for context length and model size also do not appear in any official material issued with this preview.

The doors that are actually open are Token Plan and four first-party apps

The primary path in the announcement is Token Plan, QwenCloud's credits subscription for coding and agent tools. Once subscribed, it is meant to work in tools that speak the OpenAI or Anthropic protocols; the marketing page names Qwen Code, Claude Code, Cursor, Codex, Cline, and OpenCode, among others. The personal edition is currently offered in the Singapore region. Tiers start at a low monthly price, with a limited new-customer promotion; the amount due is whatever the checkout page shows after sign-in. The plan already covered text, vision, speech, and image models. Qwen4-Max is a flagship being added to that preview roster, not a new billing product of its own.

Minutes after the post, the model list on the Token Plan page was still loading dynamically, and the public HTML did not contain the string "Qwen4." Help documentation updated on November 9 still listed qwen3.8-max as the current Qwen flagship. What developers need next is an official model ID and the credits rate, not an assumption that 3.8 calling conventions will just carry over.

Besides Token Plan, Alibaba named four first-party surfaces for early access: QwenWork, Qoder, the Qwen APP, and QwenStudio. The announcement does not say how to select the model in each product, or whether any of them has a separate quota.

The comparison that matters is still Qwen3.8-Max

Qwen3.8-Max remains the flagship in current public docs. If Qwen4-Max is going to make good on "deeper" and "stronger," it has to do it on the same class of long-horizon coding and agent work: whether a front-end comes together in one pass, whether a repo-level edit leaves existing behavior intact, and whether multi-step tool use holds up. None of that has an official number today.

Reading preview quality as the final model is also premature. A preview can change routing and reasoning policy, and it can be pulled back. The useful move is to put real problems through it, not to grade a capabilities poster.

Alibaba is putting the flagship into the tools developers already have, and leaving the spec sheet for later. That is enough to start swapping models. Aligning it with Claude, Gemini, or its own 3.8-Max still has to wait for a model ID and checks other people can rerun.

Nvdium Tech Editor