Alibaba Launches Qwen 4 Preview: Everything Known About qwen4-max-preview

On November 11, 2026, Alibaba Cloud opened limited preview access to its next-generation flagship model — **Qwen4-Max-Preview** (`qwen4-max-preview`, also aliased as `qwen4.0-max-preview`). This marks the first hands-on rollout of the Qwen 4 generation, though the full lineup and open-weights releases are still pending. **Key Specifications of qwen4-max-preview** According to Alibaba Cloud Model Studio documentation, the preview release offers: - **Modalities:** Native multimodal input supporting text, images, and video (text output). - **Context Window:** **1,000,000 tokens**. - **Maximum Output:** Up to **393,216 tokens**, dynamically shared between internal reasoning (thinking) and the final response. - **Thinking Mode:** Enabled by default, with adjustable reasoning effort (`low`, `medium`, and `xhigh` as default). - **Tools & Structured Outputs:** Built-in function calling / tool use, web search integration, plus `JSON Object` and `JSON Schema` formatting. - **API Compatibility:** Supports OpenAI Chat Completions, Responses API, and Anthropic Messages API formats. **Access Model and Preview Constraints** 1. **Token Plan Exclusive:** `qwen4-max-preview` does not support standard pay-as-you-go billing and is available exclusively through Alibaba Cloud Model Studio **Token Plan** subscriptions (Personal and Team). 2. **Built for Interactive Coding & Agents:** Token Plan endpoints (`https://token-plan.ap-southeast-1.maas.aliyuncs.com`) specifically target interactive workflows inside coding assistants and agentic tools such as **Cursor, Claude Code, Qwen Code, Qoder, and OpenClaw**. Token Plan terms prohibit automated production backends, batch pipelines, and custom scripts. 3. **Limited Testing Window:** The preview window runs from **November 11 to November 25, 2026**, after which this preview iteration is scheduled to retire. **Benchmarks and Early Developer Impressions** Alibaba **did not publish an official benchmark scorecard** alongside the initial preview launch, so formal numbers from the Qwen team are still absent. Based on early hands-on developer testing and preliminary third-party evaluations: - the model leans heavily into **agentic coding loops** (writing, checking, and iterating on code); - testers observe less overthinking on straightforward prompts compared to some late Qwen 3 releases; - early independent baseline evaluations (such as ModelDial composite scores around ~78.8/100) remain preliminary and task-dependent. **When to Expect Open Weights and the Rest of Qwen 4** Roadmap announcements mention additional tiers including **Qwen4-Plus**, **Qwen4-Flash**, and an open-weights **Qwen4-27B**, but no release date or Hugging Face / GitHub weights drop has been scheduled yet. For local deployments and stable production workloads, **Qwen 3.8** (such as `Qwen3.8-27B`) remains the active open baseline. **Bottom Line** `qwen4-max-preview` is not the general public release of the full Qwen 4 family, but a targeted flagship preview built around a 1M context window, text/image/video understanding, and deep reasoning for agentic coding. The key updates to watch after the November 25 preview cutoff will be official benchmarks and open-weights availability. [Русская версия (RU)](https://zawap.ru/posts/qwen4-max-preview-review)

Comments · 0