Qwen 4 begins: Alibaba opens Qwen4‑Max preview with a 1M‑token context

Alibaba opens Qwen 4 Max preview with 1M‑token context and multimodal I/O

November 11, 2026 — Alibaba’s Qwen team has put the first model in its next‑generation family, Qwen4‑Max, into public preview for a limited two‑week window. The preview model, exposed under the API name qwen4‑max‑preview, supports a one‑million‑token context window and accepts text, image and video inputs; it is available only to QwenCloud’s Token Plan subscribers and is scheduled to be retired when the preview ends on November 25 (UTC+8). (docs.qwencloud.com )

Tech outlets in China and internationally confirmed the launch and noted early access in Alibaba’s own apps, including the Qwen consumer app, Qoder, QwenWork and QwenStudio. (technode.com )

What Alibaba released today

• Model: Qwen4‑Max (Preview), API identifier qwen4-max-preview (alias qwen4.0-max-preview).
• Modalities: text generation with “thinking” enabled by default; vision understanding for images and video.
• Context window: 1,000,000 tokens.
• Output budget: up to 393,216 tokens, shared between internal reasoning and the final answer.
• Access: Token Plan subscription only; not available via pay‑as‑you‑go API during the preview.
• Preview window: November 11, 10:00 to November 25, 24:00 (UTC+8). The model will be unchanged during this period and then retired. (docs.qwencloud.com )

What’s still unknown

• Pricing for a pay‑as‑you‑go API call to a production Qwen4‑Max endpoint.
• Any official benchmark results, evaluation methodology, or comparative claims versus Qwen3.8‑Max.
• Parameter count and architecture details for Qwen4‑Max and the broader Qwen 4 lineup.
• Whether any Qwen 4 weights (e.g., a 27B variant) will be released openly and under what license. As of publication, Alibaba’s documentation and press materials do not disclose these items. (docs.qwencloud.com )

How to try it (developers)

Qwen4‑Max Preview is reachable only through QwenCloud’s Token Plan with a dedicated base URL. The docs note compatibility with OpenAI and Anthropic API protocols; below is a minimal OpenAI‑style call. Replace YOUR_TOKEN_PLAN_KEY with your Token Plan key.

curl https://token-plan.maas.qwencloudapi.com/v1/chat/completions \
  -H "Authorization: Bearer YOUR_TOKEN_PLAN_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen4-max-preview",
    "messages": [
      {"role": "user", "content": "Summarize this PDF and extract key risks."}
    ],
    "reasoning": {"effort": "xhigh"},
    "max_tokens": 4000
  }'

Key implementation notes from the docs:

• The preview uses a unified output budget: in OpenAI Chat Completions, max_tokens is equivalent to max_completion_tokens and covers both “thinking” and the final answer.
• Reasoning is on by default; tune with reasoning.effort (low/medium/xhigh).
• Multimodal inputs and tool calling are available; see the vision and function‑calling sections of the docs for payload shapes. (docs.qwencloud.com )

Strategic context: from Qwen 3.8 to Qwen 4

Alibaba signaled Qwen 4’s arrival in September at its Apsara Conference, saying the next‑generation model was already in training and outlining a roadmap that projects future Qwen 4.5 and Qwen 5 series at 5–10 trillion parameters. The same event highlighted a broader “full‑stack” push spanning chips, cloud infrastructure, models and agents. (alibabacloud.com )

Today’s preview follows an unusually aggressive cadence in 2026. Over the summer, Alibaba said it open‑weighted Qwen3.8‑Max with 2.4 trillion parameters, a milestone for the series and a notable marker in the open‑weights race. (data.alibabagroup.com )

Hardware backdrop: Zhenwu V900 and the scale‑up plan

At Apsara, Alibaba’s chip unit T‑Head unveiled the Zhenwu V900 accelerator and claimed a 3× performance uplift versus the prior Zhenwu M890, positioning the part for both high‑precision training and low‑precision inference. The company also set an ambition to exceed 20 GW of Alibaba Cloud data‑center capacity by 2032—an infrastructure target meant to underwrite training at trillion‑parameter scales and large‑footprint inference. (alibabacloud.com )

Early read: what Qwen4‑Max Preview changes—and what it doesn’t (yet)

• Multimodal, long‑context workflows: A 1M‑token window plus image/video understanding should reduce prompt‑engineering friction for document‑heavy and audiovisual tasks (e.g., contract review with exhibits, multi‑hour meeting analysis, and step‑by‑step report generation from long recordings). The unified output budget is a departure developers will need to account for when sizing responses. (docs.qwencloud.com )
• Enterprise agent fit: With reasoning enabled by default and tool calling built in, the preview aligns with Alibaba’s agentic cloud narrative and its QwenWork/phone‑agent push, but production SLAs, quotas and costs remain to be detailed. (alibabacloud.com )
• Evidence gap: No official scores or evaluation protocols have been published alongside the preview. For buyers and builders, that leaves comparative positioning versus Qwen3.8‑Max, Claude/Opus‑class models and other frontier releases as an open question pending formal benchmarks. (docs.qwencloud.com )

What to watch next

• From preview to GA: Whether Alibaba keeps the 1M‑token context and default reasoning behavior in the production Qwen4‑Max, and what the initial price/performance profile looks like.
• The rest of the lineup: Timing for Qwen 4 Plus, Flash and any open‑weight 27B variant, which have been widely speculated but not formally announced with specs or dates. (alibabacloud.com )
• Open‑weights policy: After open‑weighting Qwen3.8‑Max at 2.4T parameters, does Alibaba repeat that stance for any part of Qwen 4—or reserve open releases for mid‑sized models? (data.alibabagroup.com )
• Chip ramp and capacity: How quickly T‑Head’s Zhenwu V900 moves to volume and how much of Alibaba Cloud’s capacity expansion lands in AI‑optimized clusters where Qwen 4 training/inference runs. (alibabacloud.com )

Bottom line

Alibaba has started the Qwen 4 era with a controlled, time‑boxed preview rather than a full release. For developers, Qwen4‑Max Preview offers a look at the series’ long‑context, multimodal and default‑reasoning posture; for buyers, the crucial details—price, repeatable benchmarks, and model transparency—remain to come. The company’s September roadmap and hardware disclosures suggest an aggressive scale‑up path; today’s preview is the first concrete step on that path. (docs.qwencloud.com )