Alibaba Releases Qwen3.8-Max
Alibaba launches Qwen3.8 Max, a 2.4T parameter flagship AI model with 95B activated parameters, outperforming rivals in coding, research, and multimodal tasks. Weights open next week.
“AI Disruption” Publication 10,400 Subscriptions 20% Discount Offer Link.
Today, Alibaba officially released its flagship large model Qwen3.8-Max, with a total parameter count of 2.4 trillion and 95 billion activated parameters. This is currently the largest-parameter Qwen model and the first Max-tier model from Qwen planned for open weights.
Alibaba’s Hong Kong-listed shares rose 7.26% on the news, bringing its market capitalization to HK$2.41 trillion.
The upgrades in Qwen3.8-Max primarily focus on five areas: programming, office work, scientific research, long-horizon tasks, and multimodal agents. Compared with the previous-generation Qwen3.7-Max, the new model places greater emphasis on end-to-end completion of complex tasks.
Regarding API pricing, Qwen3.8-Max costs RMB 12 per million tokens for input and RMB 36 for output in China (implicit cache hit: RMB 1.5). Overseas, it is $2 per million tokens for input and $6 for output (implicit cache hit: $0.25).
On the day of release, Qwen3.8-Max also appeared near the top of three categories on the internationally authoritative Arena leaderboard.
In the Text Arena for text tasks, Qwen3.8-Max scored 1496, ranking second on the laboratory leaderboard and fifth among individual models, behind only Anthropic.
On the CodeArena WebDev leaderboard, it scored 1668, ranking third on the laboratory leaderboard and fourth among individual models, behind Claude Opus 5 Max and Kimi K3 Max.
In the Vision Arena, Qwen3.8-Max scored 1305, ranking second—behind only Claude Fable 5 and slightly ahead of Claude Opus 4.7 Thinking.
