Alibaba (Aliyun) has opened early access to Qwen4-Max Preview — the flagship model of the new Qwen4 generation, which accepts images and video as input and supports a context window of up to 1 million tokens. The preview will run until November 25, 2026: during this time, the flagship can be tried in Qwen App, Qoder, QwenWork, and QwenStudio products, as well as via API through the Token Plan subscription with a reduced credit price. The model currently has no independent benchmarks, so its quality can only be judged by its own tests for now.

image
image

What happened

On November 11, 2026, Alibaba (Aliyun) opened early access to Qwen4-Max Preview, the flagship model of the new Qwen4 generation. In the official QwenCloud documentation, the model is listed under the ID qwen4-max-preview with the alias qwen4.0-max-preview in the Token Plan subscription. The preview window is open from 11:00 on November 11 until 24:00 on November 25, 2026 (UTC+8), and for this entire period, credit consumption is reduced to 20% of the standard rate, and at night, from 22:00 to 08:00, down to 8%. According to TechNode, the model accepts images and video as input and works with a context of up to 1 million tokens. Alibaba has also integrated the new model into Qwen App, Qoder, QwenWork, and QwenStudio products, emphasizing complex programming, frontend design, and agentic tasks.

Context

Formally, this is not a final release, but a two-week preview within the Token Plan subscription, which is compatible with OpenAI and Anthropic protocols: for teams whose pipelines are already built on these APIs, the integration threshold is minimized. Judging by the announced priorities, Alibaba's bet is on the coding and agentic systems segment — complex programming, frontend design, and agentic tasks — while image and video input plus a context of up to a million tokens correspond to the set of features expected from flagship models today. A characteristic detail of the launch is that the company accompanied it with statements about improvements in knowledge, world understanding, programming, and agency, but did not publish any figures or methodology confirming these words — validation has been effectively transferred to users.

Why this matters for the industry

For the industry, the launch is primarily a price attack. Alibaba is promoting a flagship with a million-token context and video input through a promo rate, offering credit savings of up to 80%, and up to 92% during night hours, thereby putting pressure on the cost of flagship inference precisely where the battle for agentic and coding pipelines is taking place. For teams and startups, this is a two-week window of cheap experiments until November 25: you can run your own evals on the most expensive scenarios — coding on a large codebase, long agentic chains, video processing — and make migration decisions based on your own metrics, not someone else's promises. If the preview confirms the stated focus, a permanent version of Qwen4-Max with official benchmarks and standard rates is likely to follow, and another serious candidate will appear in model routing for coding and agents. A longer-term scenario — interpretation, not fact: if a million-token context and multimodal input become the standard among competitors, industry questions will shift from window size to how reasoning quality and retrieval are maintained at large lengths, and in agentic products, fewer resources will be spent on chunking and RAG, and more on orchestration, memory layers, and cost control.

Why this matters for users

You can try the model right now: via Qwen Chat, including Agent Mode, via Qwen App, Qoder, QwenWork, and QwenStudio, as well as via the Token Plan API, which understands OpenAI and Anthropic protocols. The preview window closes on November 25, 2026, so a practical plan for the remaining days is to take one or two of your most expensive pipelines — for example, working with a large codebase, a long agentic chain, or video processing — and run them at the discounted price, recording your own quality and cost metrics. Nightly batch tasks are best scheduled for the interval from 22:00 to 08:00, when credit consumption is minimal. Quality assessments at the start will have to be built solely on the results of these own tests.

What is still unknown / limitations

The main gap is quality: at the time of launch, there are no benchmarks or independent comparisons, and the company has not disclosed any figures or testing protocol. The claimed context of up to 1 million tokens requires separate verification: the sources contain neither a long-context eval protocol nor its results, and window size itself is not equal to effectively used context — it is unknown how reasoning quality and retrieval are maintained at large lengths. Statements about improvements in knowledge, world understanding, programming, and agency are not yet supported by methodology. Finally, this is a Preview: latency, service limits, and model behavior in real pipelines may differ from the announcement, so it is reasonable to postpone production decisions until the final release, for which official benchmarks and possibly a technical report are expected.

Sources

Author

Look at AI, editorial team