Chinese frontier AI models carry the fingerprint of export-controlled GPUs

A forensic analysis of publicly released model weights argues that six recent models from DeepSeek, Moonshot AI, MiniMax and Ant Group carry numerical patterns matching NVIDIA Blackwell GPUs. The method exploits how PyTorch assigns pseudorandom initial values to model weights: the arrangement depends on both the random seed and the GPU’s number of processing units. By searching for the seed and hardware layout that reproduce surviving traces in the published weights, investigators identify Blackwell configurations for DeepSeek-V4-Flash and MiniMax-M3 (148 processing units), and for DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash (152). Earlier models from these labs generally match Hopper-generation hardware. Tests against 24 models with publicly disclosed hardware reportedly agreed in every case, and the six Blackwell matches far exceed the study’s estimated chance-match threshold.

The findings suggest these labs have only recently gained access to Blackwell: DeepSeek’s shift appears between V3 and V4, Moonshot’s after Kimi K2.5, MiniMax’s after M2.7, and Ant Group’s between Ling-3.0 and 3.1. The analysis estimates final training runs required roughly 500-3,000 GPUs apiece for two to three months, implying substantial hardware costs. It does not establish where the chips were located, how they were accessed, or whether export controls were violated; the fingerprint identifies the hardware that initialized the weights, not conclusively the hardware used throughout training. Specially configured AMD systems remain a possible alternative, though investigators consider them unlikely for five models and exclude them for DeepSeek-V4-Flash.