Leaked Qwen4-Flash benchmark reportedly tops Opus 5.5
AIbase reported that leaked GameFeel Bench data placed Qwen4-Flash at 49.9, above Opus 5.5 xHigh at 47.2, while Qwen4-27B scored 44.6 against Sonnet 5.5’s 45. The same reports estimated equal-task costs of about $0.40 for Qwen4-Flash and $0.16 for Qwen4-27B, versus $2.60 for Opus 5.5. These results remain unverified, and the reports describe Qwen4 as approaching release rather than formally launched.
If independently reproduced, the reported scores and lower costs would give Qwen4-Flash a competitive claim in game-development creativity and inference efficiency, but no supplied report includes an independent run, raw outputs, or complete methodology.
11月11日,千问发布Qwen4-Max预览版,并宣布Qwen4即将正式发布。
从10月开始,关于阿里Qwen4的消息就陆续传出。此前,匿名模型Kalmia现身大模型盲测平台Arena,被广泛猜测是Qwen4;有关其游戏开发能力的内测结果,也在社交媒体上流传。
官方也在持续预热。11月6日,千问发出海报——“北冥有鱼,其名为_”。Qwen4-Max预览版官宣后,千问办公等阿里系产品第一时间接入,正式版尚未发布。
千问称,Qwen4的知识理解更深入、世界感知更立体,开发者内测结果显示,其在3D建模、网页设计、交互等方面表现亮眼。
但同期大模型领域,腾讯已经先行一步。11月4日,姚顺雨掌舵的腾讯混元发布Hy4,宣称跻身最强开源模型之列。接下来,轮到刘大一恒交出答卷了。他手中的Qwen4,能否让千问在新一轮竞争中更进一步,值得关注。
01 Qwen4想用大参数追高、小模型补缺?
11月9日,一份游戏开发创造力评估榜单GameFeel Bench在圈内热传:代号hawk的Qwen-4-Flash-Early API在内测中拿到49.9分,压过Opus 5.5 xHigh的47.2分;Qwen-4-27B-Early拿下44.6分,距Sonnet 5.5仅差0.4分。
#Qwen4-Flash#Qwen4-27B#Qwen4Plus#Qwen4Max#Opus 5.5#Sonnet 5.5#GameFeel Bench#Alibaba