🤖 NVIDIA Blackwell Fingerprint Found in DeepSeek Weights
Former OpenAI researcher Will DePue found a fingerprint of export-controlled NVIDIA Blackwell GPUs in the public weights of six models from DeepSeek, Moonshot AI, MiniMax, and Ant Group in an Unemployment Labs report. The method reconstructs the seed and PyTorch thread layout (PRNG Philox), which depends on the number of SMs in the GPU. On a real B300, DePue reproduced 445 out of 469 lines of DeepSeek-V4-Flash weights bit-for-bit; there were no matches on H100.
🌍 The report estimates the scale: five final training runs are equivalent to ~12,000 Blackwell GPUs and ~30,000 H100s, and the cluster for Kimi K3 cost $200–300 million. This increases pressure on US export controls.
👤 Public weights can be used to determine on what hardware a model was initialized. Alibaba, Zhipu, Tencent, Xiaomi, and StepFun currently remain on Hopper, while Meituan trains on its own ASICs.
Source 1: https://unemploymentlabs.com/checkpoint-forensics