
What happened
Alibaba launched Qwen4-Max, built on a new Qwen4 architecture, with a 2.4 trillion parameter backbone that activates only 71 billion at a time and adds 328 billion N-gram parameters as a cache.
Why it matters
The Qwen4 architecture makes training more efficient, and the upgraded attention mechanism lowers infrastructure costs, so the model costs less to run while producing higher-quality output than its predecessor, according to Alibaba.
WHO IT HITSEnterprise customers using Alibaba's cloud platform can access Qwen4-Max, and developers building long-horizon tasks may benefit from the open-sourced Qwen-Live-Delivery-Plugin that shortens wait times by delivering iterative drafts.
Summaries like this, every morning.
Alibaba's Qwen4-Max arrives as an upgrade to its Qwen3.8-Max predecessor, with the company claiming both higher output quality and lower cost. The key architectural change is the Qwen4 backbone, which activates only a fraction of its 2.4 trillion parameters at any given time and relies on 328 billion N-gram parameters as a cache to avoid recomputing frequent calculations. This design, along with a four-channel residual mechanism that lets gradients skip layers during training, appears aimed at making both training and inference more efficient.
The attention mechanism has also been customized: Gated DeltaNet reduces hardware needs for large prompts, while QSA skips unnecessary data. Alibaba reports that on an internal benchmark of knowledge work tasks, Qwen4-Max used up to 35% fewer tokens than Qwen3.8-Max, and savings reached 49% on the Terminal-Bench 4.0 coding suite. These figures suggest the model could be cheaper to operate for customers, though independent verification is not yet available.
Beyond the core model, Alibaba is open-sourcing a plugin that lets Qwen4-Max deliver an initial draft quickly and then iteratively refine it with user feedback, shortening waits on long-horizon projects. The company also plans to open-source smaller models based on the same technology, which could broaden adoption among developers who need lighter-weight options.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Noeri Inc. launched with $110 million in funding led by Menlo Ventures, with Madrona Ventures, Kindred Venture…
A Qiita tutorial from engineer skill-sharing platform DokuPro walks through collecting thumbs-up/thumbs-down r…

Duco moved its agentic AI stack to Amazon Bedrock and AgentCore, reducing main context from 200,000 tokens to…

Fierce Network reports that AI agents are operating in production environments, with the implication that they…

CrowdStrike was named the Visionary Leader in the Frost Radar™: AI Security Operations Centers, 2026, position…

Verizon CEO Dan Schulman said AI in customer service delivers a 1,200-basis-point or more improvement in custo…
