According to reports on November 8th, as a new round of domestic AI large models enter the countdown for release, the core performance scores of Alibaba’s Qianwen Qwen4 series have been revealed earlier. In a special creativity evaluation focused on game development, some models in this series successfully surpassed the top overseas model Opus5.5, demonstrating a remarkable pace of technological advancement.
Game development creativity benchmark score revealed: Flash successfully surpasses Opus5.5
The revealed evaluation benchmark is not a traditional general large model score, but rather a creativity assessment project in the game development field called GameFeel Bench.
The test results show that the lightweight Qwen4-Flash achieved a score of 49.9, successfully surpassing Opus5.5xHigh, which got 47.2 points. This is the first time a domestic large model has surpassed a foreign competitor at this level in relevant benchmark tests. Meanwhile, the smaller local deployment version of Qwen4-27B also received a high score of 44.6, closely approaching Sonnet5.5’s 45-point score.
Exquisite cost-performance ratio and a rich product portfolio
In addition to achieving breakthroughs in performance, the Qwen4 series also has significant advantages in inference usage costs.
According to the leaked information, the cost for Qwen4-Flash to complete the same task is only about $0.4, far lower than the $2.6 required by Opus5.5; while the execution cost of Qwen4-27B has been reduced to $0.16, making it highly cost-competitive among products of the same class.
In terms of product layout, the Qwen4 family includes at least 4 products with different positioning. In addition to the Flash and 27B versions that stood out in the benchmarks, there is also Qwen4Plus, as well as Qwen4Max, which represents the highest performance limit of this series. As the release date promised by the relevant parties approaches, the official launch of these models will provide the industry with a new window for direct testing and comparison with overseas flagship models such as GPT-6.1 and Opus5.5.