In the upcoming new wave of domestically developed Chinese AI large models, Alibaba's Qwen4 series is the first to show strong technical momentum. According to recently leaked core benchmark data, the series performed impressively on GameFeel Bench, a benchmark specifically designed to evaluate game-development creativity, successfully surpassing the top overseas model Opus5.5 on some metrics and becoming the first domestic large model to overtake it in this benchmark.
Core scores beat overseas rival as local deployment version closely trails mainstream
The leaked test results show that the lightweight Qwen4-Flash achieved an excellent score of 49.9, officially surpassing Opus5.5xHigh, which scored 47.2, marking a new height for domestic models in creative performance in a specific vertical domain. At the same time, the smaller-parameter local deployment version Qwen4-27B also scored a high 44.6, putting it almost on par with Sonnet5.5, which scored 45.
In addition to its performance breakthrough, the series also has a significant advantage in inference cost. Data shows that Qwen4-Flash costs only about $0.4 to complete an equivalent task, far lower than the $2.6 required by Opus5.5; Qwen4-27B's execution cost is compressed to just $0.16, demonstrating extremely strong cost competitiveness among products in the same tier.
A complete product matrix is about to debut, with multiple flagship models poised for launch
In terms of product lineup, the Qwen4 family has at least four core models with different positioning. In addition to the Flash and 27B versions that have already made their mark in benchmark scores, it also includes Qwen4Plus, aimed at broader application scenarios, and Qwen4Max, which represents the series' highest performance ceiling. As the release window previously promised by the Qwen team approaches, the official launch of the series will provide the industry with a new window to observe domestic large models competing head-to-head with overseas flagships such as GPT-6.1 and Opus5.5.
