AboutAI ToolsCursorClaude CodeLovableBolt.newAI tool reviewsTutorialsVibe codingAI agentsMobile apps with AIAll tutorialsBlogSponsorWork With MeJoin 285K+ Subscribers

11 November 2026 · 3 min read

Alibaba Opens Qwen4-Max Preview, Targeting Complex Coding and Agentic Workflows

The Qwen4-Max preview is now available across Alibaba’s products, with a focus on complex programming, front-end design, and autonomous tool use. Pricing includes promotional discounts and a ¥690/month Qwen Studio plan.

Alibaba has begun rolling out early access to its next-generation Qwen4 model family, with the flagship Qwen4-Max preview leading the release on 11 November 2026 [2][4]. The model is optimised for complex coding, front-end design and agentic tasks, with a particular focus on professional office work and software operations — enabling it to autonomously call tools and complete multi-step tasks [1][2].

What is new

Rather than a single entry point, the Qwen4-Max preview is being integrated simultaneously into Alibaba’s Token Plan, Qwen Office, Qoder, the Qwen App and Qwen Studio [1][4]. This allows developers and enterprise users to trial the model inside real workflows before an official launch.

Early tests suggest several practical strengths:

  • In a web development task, Qwen4-Max reportedly outperformed Deepseek 4.1 Flash while using roughly half the tokens [3].
  • It generated a detailed Minecraft clone with shaders, day-night cycles, water dynamics and multiple terrains, though minor flaws such as an inability to exit water were noted [3].
  • A Roman coliseum scene was produced in “high thinking” mode consuming about 69,000 tokens, showing strong composition without needing the highest reasoning setting [3].
  • On simulation tasks, a Hermes-powered version simulated a bioreactor producing insulin, with adjustable parameters and predicted yields. The results are not validated for real-world use, but they indicate potential for scientific and industrial applications [3].
  • The model generated a Call of Duty-style game with mechanics and NPCs, and created a 3D boat environment using Three.js, outperforming GPT-6 Astra with fewer tokens [3].

A promotional image leaked by an Alibaba partner suggested the Max variant may include a 1-million-token context window, image support and tool-calling capabilities, though these specifications have not been officially confirmed [3].

Alibaba has not yet disclosed Qwen4-Max’s parameter scale, benchmark scores or an official launch timeline [1]. By comparison, the current Qwen3.8-Max uses a mixture-of-experts architecture with 2.4 trillion total parameters and 95 billion active parameters [1].

Availability and pricing

  • Token Plan: A flat-rate subscription under QwenCloud. During the promotional period, new users can subscribe for as little as $3 per month, with a limited-time discount of up to 90% [1]. A single contract provides access to multiple models beyond Qwen and allows users to designate AI tools that support OpenAI and Anthropic protocols, such as Claude Code, Cursor and Qwen Code [1].
  • Qwen Studio: Offers a free plan and three paid tiers — Lite (¥690/month, roughly $4.3), Plus (¥2,890/month) and Pro (¥16,800/month). The plans differ in usage quotas for models and agentic features; Qwen Code quota under Plus and Pro is shared with agentic capabilities [1].

Later in the Qwen4 series, Alibaba plans to add Flash and Plus versions, as well as a 27-billion-parameter model for on-premises deployment [1][2].

What it means for developers

The decision to launch a preview across multiple product lines — rather than waiting for a polished release — signals that Alibaba wants to lock in early developer mindshare. Developers can test Qwen4-Max inside familiar tools such as the Qwen App and Qwen Studio, while the Token Plan’s flat-rate pricing could make it easier to budget for API costs [1].

For those building agentic applications or tools that require reliable multi-step software operations, Qwen4-Max’s emphasis on autonomous tool calling is a key point of evaluation. Early test outputs, though not benchmark-grade, suggest the model handles token-intensive creative and simulation tasks efficiently [3].

Rumours that Alibaba might abandon open-weight development for smaller models have been denied. Slides from the Esper conference confirmed that both Qwen4 and the future Qwen5 will remain open-weight [3]. The previously released Qwen3.8-Flash-Next already gave a glimpse of the new architecture — combining compressed context memory with key-passage rereading — and the 27B on-premises version of Qwen4 is expected to build on that direction [1][2].

Until official benchmark results and parameter counts appear, the real question for developers is whether the preview holds up in demanding production settings. The promotional pricing lowers the barrier to trying it out, but actual performance against next-generation competitors is still unverified.

Sources

  1. Alibaba's Qwen4-Max Preview Launches Abruptly, Targeting Complex Coding and Professional Software Operations — BigGo Finance
  2. Aliyun Launches Qwen4-Max Preview Focused on Complex Programming | KuCoin
  3. Qwen 4 Beats Deepseek 4.1 Flash in Web Tasks on Half the Tokens
  4. Qwen4-Max预览版已首发上线阿里Token Plan
  • qwen4
  • alibaba
  • coding-model
  • agentic-ai
  • developer-tools

Researched and drafted with AI from the sources above.

Learn to build with AI

Hands-on tutorials on Cursor, Claude Code, AI agents and more.