Newshot

Alibaba's 2.4-Trillion-Parameter Qwen4-Max Officially Launches: Task Costs Drop Over 60%, Smaller Open-Source Models to Follow

AI summary

On November 16, IT Home reported that Alibaba's Qwen team announced the new flagship Qwen4-Max, built on a sparse MoE architecture with 2.4 trillion main parameters, native visual understanding, and a 1 million-token context window; it is now officially available on the Qwen AI platform via Token Plan or API, following a limited preview on November 11. Qwen says the model improves complex work tasks while reducing token consumption and task costs: it scores 52.69 on Work360 versus 47.74 for Qwen3.8-Max, with token use at Xhigh and High reasoning intensities down about 24% and 35%, respectively, and down about 48% and 49% on Terminal-Bench 4.0. API prices have also been lowered, especially cache-hit pricing, with the average cost per task in Work360 evaluations falling from 26.58 yuan to 11.16 yuan, a 58% drop; the team also developed the Work360 evaluation framework to assess real-world professional work scenarios.

Translation(model translation, not the source's own text)

Alibaba's Qwen4-Max with 2.4 Trillion Parameters Officially Launched: Task Costs Cut by Over 60%, Smaller Models to Be Open-Sourced Later

IT Home News, November 16: The Alibaba Qwen team today announced the launch of its new-generation flagship model Qwen4-Max, featuring an all-new MoE (Mixture of Experts) architecture with a backbone parameter scale of 2.4 trillion, native support for visual understanding, and a 1 million token context window.

Unlike the limited-time preview version that was first opened on November 11, this release marks Qwen4-Max's official integration into the Qwen AI platform as a formal version, with developers able to access it via Token Plan or API calls.

Qwen4-Max adopts a sparse MoE architecture, with systematic adjustments made across three levels: Attention, Residual, and Embedding. Three out of every four layers use Gated DeltaNet to compress historical information into a fixed state, while the remaining layer uses Qwen Sparse Attention to select relevant context at block granularity, reducing the attention computation overhead for long sequences; Gated Residual expands the residual stream into four parallel branches to enhance cross-layer information transfer; and N-gram Embedding provides local token pattern representations via table lookup at the second layer, expanding model capacity with relatively low additional compute.

According to the official statement, the new model, while improving its ability to handle complex work tasks, can reduce token consumption and lower task costs.

According to the official Qwen blog, Qwen4-Max targets scenarios such as office productivity, creative work, software development, and scientific research, with a focus on improving professional tool use capabilities, long-cycle task execution capabilities, and final delivery quality. The model can handle tasks in 3D modeling, game development, industrial manufacturing, finance, law, healthcare, and frontier scientific research.

Efficiency is the headline metric of this upgrade. Qwen4-Max scored 52.69 on the Work360 industry benchmark, compared with 47.74 for the previous generation Qwen3.8-Max.

At equivalent performance, the token consumption at the Xhigh and High reasoning intensity levels was reduced by approximately 24% and 35% respectively compared with the previous generation; on Terminal-Bench 4.0, the token consumption reductions reached approximately 48% and 49%. API pricing was also lowered accordingly, with the cache-hit price in particular cut substantially—the average cost per task in the Work360 evaluation dropped from 26.58 yuan to 11.16 yuan, a reduction of 58%.

To evaluate the model's performance in real work scenarios, the Qwen team also developed the Work360 evaluation framework to simulate and classify complex professional work tasks. IT Home notes that the framework covers 26 primary domains, 182 secondary subdomains, and 721 tertiary professional directions, spanning workflows across multiple industries and occupations.

The model's capabilities lean toward long-horizon task execution. Qwen4-Max supports a million-token context window and native visual input, and introduces a Live Delivery mechanism in agent scenarios: first deliver a usable initial version, then iteratively expand content and quality based on user feedback, with the Qwen-Live-Delivery-Plugin released as open source alongside it.

The officially showcased cases cover Blender 3D modeling, game development with the Unity and Unreal engines, automotive production line planning and parts machining programming, as well as professional scenarios such as financial investment research, legal due diligence, and critical care clinical assistance. In the automotive parts machining case, the manual effort for programming a single workpiece dropped from about 475 minutes to 24 minutes.

Image
Image

Qwen4-Max is the first model in the Qwen4 series. According to the official statement, the smaller models in the series will be open-sourced successively, following the same technical lineage while making trade-offs for scenarios where smaller models excel.

The Qwen4-Max API is now live on QwenCloud, and developers can access it via API. In addition, users can experience the new model through Qwen Studio, QwenWork, and Qoder.

  • Blog: https://qwen.ai/blog?id=qwen4-max
  • API: https://www.qianwenai.com/models/qwen4-max
  • Token Plan: https://platform.qianwenai.com/pricing/token-plan

Related Articles

  • MIT researchers propose "weight-space programming": GPT-6 Astra modifies only about 0.016% of Qwen3.5-0.8B's weights, boosting math scores to 6 times the original
  • Cook: Apple Intelligence to launch test version in China by end of month, but without Siri AI
  • Alibaba Cloud's Brazil node opens, Qwen series open-source models surpass 3 billion global downloads
  • Alibaba Qwen3.8-Flash (Next) released as open source: 125B-parameter MoE model with training cost only 1/9 of previous generation
  • World's first: Alibaba Qwen AI models surpass 3 billion global downloads, crushing Meta and Google
  • Report: Alibaba plans to charge revenue share from large commercial users of next-generation Qwen open-source models

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗