Researcher finds numerical traces of Nvidia Blackwell in six Chinese models
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
A model's published weights can reveal the generation of hardware used to create it, providing a new technical way to inspect training infrastructure without access to a data center.
Former OpenAI researcher Will DePue reported that the published weights of DeepSeek-V4-Flash and V4-Pro, Kimi K3 and K3.1-Pro, MiniMax-M3 and Ling-3.1-flash retained characteristic initialization patterns from Blackwell GPUs. The method uses differences in how random starting values are distributed across compute units: the B200/B300 have 148 SM units, while the GB200/GB300 have 152. The B300 445-of-469 check reproduced selected DeepSeek-V4-Flash weight rows bit for bit. The GB200 was used for Kimi K3. The method also correctly identified the hardware of all 24 control models with known equipment. The authors estimate that the five largest runs may have required about 12 thousand Blackwell GPUs, but this is a calculation rather than a confirmed figure. The analysis shows which hardware initialized the weights, but does not establish the GPUs' location, owner, means of access or any violation of export restrictions.
Sources 2