Researchers link Chinese AI model weights to Nvidia Blackwell hardware

UnemploymentLabs researchers say statistical patterns in the publicly available weights of DeepSeek V4, Kimi K3.1, MiniMax-M3 and Ling-3.1-Flash point to training on Nvidia Blackwell hardware, Tom's Hardware reports. The analysis does not establish where the chips were located, how developers accessed them or whether export controls were violated. The method reproduces how training software initializes model weights across accelerator processing units, then compares the resulting distributions. Changing processing-unit and thread counts broke the correlation in tests. The researchers also ruled out AMD hardware for DeepSeek V4 by examining weight values. Their technique applies only to models with open weights and cannot distinguish exact card variants from initialization patterns alone.