header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP
Partnered With:

Qwen4-Max Officially Released: Backbone Scale Unchanged, Activated Parameters Reduced by 25%, Task Cost Down 62%

Beating AI News Flash: Following last week's open preview, Alibaba has officially released its new-generation flagship model Qwen4-Max, disclosing the full technical architecture, benchmark results, and API pricing.


Qwen4-Max retains the 2.4 trillion parameter backbone scale of the previous generation and additionally introduces 328 billion N-gram Embedding parameters. By redesigning the attention mechanism, residual connections, and phrase embedding layers, the backbone parameters activated per Token generation have been reduced from approximately 95 billion to 71 billion, a decrease of about 25%. The model supports native visual understanding and a 1 million Token context.


Official benchmarks show that Qwen4-Max scored 52.69 on the professional work benchmark Work360, up from 47.74 for the previous generation. This Alibaba-built evaluation covers professional tasks in finance, law, manufacturing, and other fields. In the Terminal-Bench 4.0 coding test, the new model's Token consumption was reduced by about 48% to 49% compared with the previous generation. The average per-task cost on Work360 dropped from $4.43 to $1.67, a decline of 62%.


Alibaba also presented a scientific research case. Qwen4, together with an automated experimentation platform, ran autonomously for 48 consecutive hours, screening 2.7 million candidate molecules for new cooling fluid materials suitable for data centers. After four rounds of molecular design, experimental synthesis, and testing, a physically validated candidate molecule sample was ultimately produced. However, officials have not yet disclosed data sufficient to prove that its actual cooling performance is superior to existing materials.


In an automotive manufacturing case, Qwen4-Max participated in part modeling, machining path planning, and simulation verification. For one workpiece machining programming task, manual operation time was shortened from about 475 minutes to 24 minutes, a reduction of about 95%. The model can also use Blender, Unreal Engine, and Unity to create 3D scenes and playable games, and iteratively revise them based on runtime feedback.


Qwen4-Max has opened its official API, supporting OpenAI- and Anthropic-compatible interfaces. Pricing is 10 yuan per million input Tokens and 30 yuan per million output Tokens, about 17% cheaper than the previous generation; the cached-hit input price has dropped from 1.5 yuan to 0.5 yuan.


Alibaba confirmed that smaller models in the Qwen4 series will be open-sourced later.

举报 Correction/Report
Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish