Alibaba on November 16, 2026 released Qwen4-Max, which it calls the most capable model in the Qwen family to date and the first model of its Qwen4 generation. In its announcement, the company describes a mixture-of-experts model with 2.4 trillion total parameters, native vision capabilities and a one-million-token context window, aimed at coding, agentic tasks, office productivity and long-running research work.

Because it is a mixture-of-experts design, only part of the network runs for each token. Alibaba says Qwen4-Max activates 71 billion parameters per token, 25% fewer than Qwen3.8-Max at the same 2.4 trillion scale. The company attributes its gains to a new Qwen4 architecture that upgrades the sparse design across attention, residual connections and embeddings. MarkTechPost's technical breakdown describes those pieces as Qwen Sparse Attention layers interleaved with Gated DeltaNet layers, gated residual branches and an additional 328 billion N-gram embedding parameters.

On performance, Alibaba says the model ranked sixth on Code Arena, a real-world coding leaderboard, and posted strong scores on agent, workplace and visual-reasoning benchmarks including Agents' Last Exam, Vals Index, JobBench and Chartography. On Work360, a proprietary Alibaba benchmark spanning 721 specialized fields, the company says Qwen4-Max beats Qwen3.8-Max overall while using 35% fewer tokens at its high reasoning-effort setting and cutting average cost per task by around 60%. The Next Web reports the Work360 average task cost fell from $4.43 to $1.67. MarkTechPost characterizes the model as competitive with frontier systems on agentic work but behind on the hardest coding and vision tests.

Qwen4-Max is available now through the QwenCloud API and Token Plan and in Alibaba apps including QwenWork and Qwen Studio. Alibaba Cloud's Model Studio documentation lists the model ID qwen4-max with compatibility for the OpenAI and Anthropic APIs, support for web search, tool calling and structured outputs, and thinking mode enabled by default. The Next Web reports API pricing of $1.50 per million input tokens and $4.50 per million output tokens, and Alibaba's release lists cached input at $0.075 per million tokens. According to Reuters, as carried by Lianhe Zaobao, Alibaba also said it would cut API prices across the board.

The release departs from much of Qwen's history in one notable way: Qwen4-Max's weights are closed. Alibaba says smaller Qwen4 models will be released with open weights soon, without giving a date. The Next Web notes that Qwen is already the most downloaded open-weight model family in the world, according to Bloomberg, which makes the decision to keep the flagship hosted-only a meaningful shift for developers who self-host.

Alibaba also highlighted long-running agent demonstrations, including a continuous 48-hour closed-loop run that it says discovered an experimentally validated coolant molecule for chip cooling, and a manufacturing showcase in which manual workpiece programming time fell from roughly 475 minutes to 24. On safety, the company says it filtered training data to keep defensive security material while removing offensive attack content, added code-security checks to software training, and strengthened protections against chemical, biological, radiological and nuclear risks.

Readers should weigh the caveats. The benchmark rankings, cost savings and agent demonstrations are Alibaba's own claims, Work360 is a proprietary test, and independent evaluations had not been published at launch. DigiEditorial's account is based on Alibaba Cloud's announcement and Model Studio documentation, corroborated by The Next Web, MarkTechPost and Reuters reporting carried by Lianhe Zaobao.