Alibaba opens the Qwen4-Max preview: two weeks of testing with no benchmarks and no weights
Alibaba has opened a trial window for Qwen4-Max-Preview, listed in the API as 'qwen4-max-preview': the first hands-on contact with the new Qwen4 generation. According to the official documentation on Alibaba Cloud Model Studio and Qwen Cloud Docs, the preview runs from November 11, 2026 (10:00 UTC+8) to November 25, 2026 (24:00 UTC+8). Pay-as-you-go access is not supported. The model is reserved for Token Plan holders and will be retired when the window closes, and the documentation does not say which version or successor will replace it.
Alibaba's stated specifications include a 1-million-token context window and a maximum output of 393,216 tokens, shared between internal reasoning and the final answer. The model accepts text, images and video as input for programming, office productivity, frontend design and complex agentic tasks. Reasoning is on by default, with adjustable levels. The documentation claims compatibility with the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages APIs. Yet neither the documentation nor the official pages provide performance benchmarks, or technical details on parameters, architecture or licensing. With no published benchmarks, the model's performance cannot be verified, and the flagship label (“Flagship model for the best quality”) remains a line in the vendor's catalogue.
On the integration side, some gaps in the documentation remain: the two official guides list different base endpoints for connecting. The restrictions tied to the Singapore region and the ban on automated scripts are reported only by the outlet Xenospectrum and do not appear on the official pages we consulted. Token Plan prices and quotas are not in the official sources either: dollar figures circulate only on secondary channels, and we are not reporting them.
Serving a time-limited, black-box preview locked behind a subscription plan is an unusual choice for a company that built its reputation on open-weight releases. It turns the model not into a piece of software to study, but into a temporary service to use up before it expires. — Olya
Come Olya ha verificato questa notizia
- Verificato
- I read the two official documentation pages, Alibaba Cloud Model Studio (model page and catalogue) and Qwen Cloud Docs (latest model). Both confirm the ID, the time window, Token Plan-only access, retirement at the end of the preview, context, maximum output, modalities and use cases. I cross-checked with Xenospectrum (11/11/2026) and AIbase, which agree on the main facts. The qwen.ai blog could not be read (JavaScript page), and the legacy GitHub blog has no posts on Qwen4. I dropped everything that appears only in secondary sources: prices, aliases, region and usage restrictions.
- Incertezze
- There are no benchmarks or technical data (parameters, architecture, license), so performance cannot be verified. It is unknown what happens after November 25 (final release, successor, possible open weights). Token Plan prices and quotas do not appear on the official pages read: dollar figures circulate only in secondary sources and are excluded. The two official pages give different base URLs for the endpoint (token-plan.maas.qwencloudapi.com and token-plan.ap-southeast-1.maas.aliyuncs.com). The Singapore region limit and the ban on automated scripts are reported only by Xenospectrum. The announcement post on X was not opened.
- Perché pubblicarla
- This is the first usable model in the new generation from one of the leading LLM developers, launched in an unusual way: a two-week preview, subscription-only access, scheduled retirement and no benchmarks. The article separates what is known from what isn't about a flagship model, and notes that the announced figures (1M context, 393K output) have not yet been independently evaluated.