Alibaba has opened a preview of Qwen4-Max, the first model from its next flagship Qwen4 family, aimed at software development and long-running AI agent tasks. The preview went live on 10 November and was reported on 11 November 2026.

Alibaba says Qwen4-Max improves on its predecessor, Qwen3.8, in world knowledge, code generation and sustained performance on long tasks that need repeated tool calls. Those claims cannot yet be checked independently, because Alibaba has not published benchmarks or a model card for Qwen4-Max.

How developers can try Qwen4-Max

Developers can reach the preview through Alibaba's Token Plan, a subscription that bundles model use for coding agents and other API clients, with advertised entry pricing starting at about $3. Alibaba is also rolling out early access in QwenWork for workspace automation, Qoder for agent-assisted software development, the Qwen App for mobile access, and QwenStudio for custom agent development.

The preview also works with clients that use OpenAI- or Anthropic-compatible protocols, including Qwen Code, Cline, Claude Code, Cursor, OpenCode and Codex. That means teams already using one of those tools can try the model by changing the provider settings rather than rebuilding their setup, although feature support can still vary between clients.

What is still unknown about Qwen4-Max

Alibaba has not published a parameter count, benchmark results, pricing for pay-as-you-go use, or a date for general availability. It has also not announced downloadable weights or self-hosting terms for Qwen4-Max, so for now the model is available only as a managed preview.

Tracker BenchLM records the preview as qwen4-max-preview on Alibaba's Token Plan since 10 November, and notes that Qwen Chat lists the model with a 1M-token chat context, thinking, and image, video, audio and document input. That listing is not an official model specification, and client tooling that assigns a one-million-token allocation to the model identifier does not by itself confirm the usable context window.

The road to Qwen4

The preview follows weeks of signals from Alibaba. At its Apsara Conference in Amsterdam on 29 October, a presentation listed Qwen4-Max, Qwen4-Flash, Qwen4-Plus and Qwen4-27B as forthcoming models, and Alibaba Cloud chief technology officer Feifei Li reportedly told attendees the family would arrive the following week. A model known as Qoduck later appeared in Alibaba's Qoder coding tool, and Chinese outlet 53AI identified it as an early Qwen4-Max checkpoint, although Alibaba did not confirm that identification.

Alibaba said in September that Qwen4 was in training, with Qwen4.5 and Qwen5 planned to scale to 5 to 10 trillion parameters. For now, the open question for developers is how Qwen4-Max performs in real coding work, and that will only become clear once independent tests and benchmarks appear.

Sources: AlphaSignal and BenchLM.