Research result · Sep 12, 2026
Caiyun Technology team twice breaks the NanoGPT Speedrun large-model training speed world record, code merged into the official open-source repo
Caiyun Technology announced that its base-model algorithm team has twice set a new world record in the open-source NanoGPT Speedrun training-speed benchmark, cutting time-to-target to 1 minute 16 seconds. Its code has been merged into the official open-source repository.
The two records rest on two original model architectures: DCMHA (dynamically composable multi-head attention) and MUDD (multi-path dynamic dense connections). Both were published at ICML as oral presentations, in 2024 and 2025 respectively.
Original sources (Chinese)
9点1氪丨房贷延长到40年但申请人不能超过35岁？机构回应；长鑫利润率反超三星SK海力士；戴尔市值一夜飙升2500亿元，创历史新高