Qwen4-Max preview opens with 1M context and a two-week test window

Alibaba has opened a Qwen4-Max preview through Token Plan, with a 1M-token context window. The documented preview runs from November 11, 2026 at 02:00 UTC until November 25 at 16:00 UTC, after which the model retires. Those times convert Alibaba’s UTC+8 schedule, including its end-of-day deadline.

For developers, the useful question is what to learn before committing a workflow to a new model. A larger context budget can accommodate more material, but the result still needs to be correct, reviewable and worth the time spent waiting. ByteForward has reviewed the access documentation; we have not benchmarked the preview.

What is available today

The Model Studio catalog lists qwen4-max-preview as Token Plan-only, accepting text, images and video and returning text. It supports thinking, function calling, built-in tools and structured output. Its maximum 393,216 tokens combine thinking and output, rather than promising that many tokens of visible answer.

That is firmer evidence than the Qwen Code preparation changes we examined earlier. That report covered client defaults before documented access. Today’s guide concerns this specific hosted preview; the reviewed pages do not establish a complete Qwen4 family release or downloadable Qwen4 weights.

Treat the context and output figures as separate planning inputs. A repository task may spend most of its budget reading code, while a long explanation spends more on generation. Neither ceiling tells you whether the model found the relevant dependency or preserved an important constraint halfway through a session.

Check the plan before configuring the client

The Personal Edition overview specifies Singapore with Global inference and cross-border data transfer. The model also accepts qwen4.0-max-preview; billing uses that alias. Promotional consumption is advertised as low as 20% of the standard rate, or 8% between 22:00 and 08:00 UTC+8. Actual Credits depend on usage.

The plan FAQ restricts Personal Edition to individual interactive development in supported tools. Production backends, batch scripts and scheduled non-interactive work are prohibited, and accounts cannot be shared. The plan itself has no free trial quota. Monthly Credits reset on the subscription anniversary, not a common calendar date; calls stop after plan and any purchased bundle Credits run out, rather than automatically becoming pay-as-you-go.

Before sending company code or documents, check the organization’s approved processing locations and vendor arrangements. A region selector alone is insufficient evidence that a workload meets an internal residency requirement. For an evaluation, start with a small public repository or material you are explicitly allowed to send.

Use the matching key and endpoint

Alibaba’s quick-start guide separates Token Plan credentials from standard Model Studio and legacy Coding Plan credentials. Token Plan keys begin with sk-sp- and must be paired with the corresponding endpoint. OpenAI-compatible clients use https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1. Anthropic-compatible clients use https://token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic.

For Qwen Code, the official connection guide starts with selecting Alibaba ModelStudio, the billing plan and Singapore, then entering the matching key. Its desktop flow lets users add models in Settings and select the saved model. Check the exact preview identifier rather than assuming that an automatic router chose it. Keep credentials out of repository files, screenshots and shared transcripts.

Make one small interactive request first, then inspect the provider’s consumption record before expanding the test. The FAQ warns that incorrect key and endpoint configuration can route calls to separate pay-as-you-go billing. It also notes that tool schemas, conversation history and reasoning contribute to consumption beyond the tokens a third-party client may display.

A useful evaluation should expose failure

Choose a task with a result you can independently check. For coding, use a bounded change with existing tests and a known expected behavior. Save the starting revision, ask for a plan, review the diff and run the relevant checks. A persuasive explanation is not a substitute for a working change.

Keep the client version, tools, permissions and reasoning settings consistent when comparing models. Record whether the task completed, the amount of manual repair, elapsed time and observed cost. Include an unsuccessful attempt in the notes. Otherwise the comparison can quietly turn into a showcase of whichever model got the easiest prompt.

For long-context work, place essential constraints in different parts of the input and require answers tied to specific files or passages. For visual work, use an image or short clip whose important details you already know. Check both missed details and confident inventions. These are suggested evaluation methods, not results we observed from Qwen4-Max.

Leave yourself an exit path

Store useful task summaries and accepted changes outside the conversation. Keep a second supported model configuration ready, and test whether it can continue from a concise handover. That is basic operational hygiene for any time-limited preview, especially when a demo could become part of someone’s daily routine.

The decision to keep using a model should follow the evidence from your own work: fewer corrections, acceptable latency and a clear cost record. The new specification provides a reason to evaluate Qwen4-Max. It does not settle those questions on its own.