Primary source掘金本周最热
View sourceEvent ·Multi-source
腾讯混元 Hy4 正式版发布:加视觉与 Computer Use,幻觉错误率从 23.1% 降到 5.6%
11 月 4 日腾讯发布 Hunyuan Hy4 正式版,底座与 8 月开源的 Hy4 preview 相同,仍是 770B 总参数、49B 激活的 MoE。正式版新增视觉理解和 CUA 计算机操作能力,在 OSWorld 2.1 拿 78.9 分,排 Kimi K3.1 Pro 和 Claude Opus 5.5 之后。因猜测导致的错误率从 23.1% 降到 5.6%。官方十二格评测看板对 Claude Opus 5.5 十二项全负,八项排第三;第三方 SuperCLUE 中文榜上腾讯目前只有 Hy3 的 62.13 分在列。
Current read
腾讯混元 Hy4 正式版上线,这版加上了看图和自己操作电脑的能力,砍幻觉的数字也给得很具体,跟 DeepSeek、Kimi、GLM 的位次文中都摆出来了。
11 sources71° AI heatLast updated 2026/11/4 08:14:16
Evidence
11 sourcesRelated source 1shao__meng
View sourceRelated source 2Artificial Analysis
View sourceRelated source 3lmarena.ai
View sourceRelated source 4Hunyuan
View sourceRelated source 5Nous Research
View sourceRelated source 6Pandaily
View sourceRelated source 7TestingCatalog
View sourceRelated source 8kimmonismus
View sourceRelated source 9腾讯混元
View sourceRelated source 10Nikkei Asia
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.