A new technical report claims that the publicly released weights of several Chinese frontier AI models carry a fingerprint of the Nvidia GPUs used to create them, and that six of those models point to Blackwell-generation hardware, which US export controls restrict. The report, "Checkpoint Forensics," was published on November 12 by Will DePue of Unemployment Labs. It is not peer reviewed.
The method relies on how PyTorch generates random starting weights. According to the report, PyTorch's GPU random number generation splits work across a chip's streaming multiprocessors (SMs), so the SM count leaves a layout pattern in the initial weights. By searching combinations of random seeds and layouts, the author says he can recover that pattern, either directly from rows that were never trained, such as reserved vocabulary entries, or from a weaker correlation that survives in trained weights. He reproduced specific rows with unmodified PyTorch on rented H100, B300 and GB200 hardware, and tested seeds chosen on one set of weights against separate weights.
The report says six models show Blackwell-compatible SM counts: DeepSeek-V4-Flash (148 SMs), DeepSeek-V4-Pro (152), Kimi K3 and K3.1-Pro (152), MiniMax-M3 (148) and Ling-3.1-flash (152). It calls the evidence for DeepSeek-V4-Flash "conclusive" and reports statistical strength between 49 and 1,385 sigma across the six, against a chance threshold of about 8 sigma. Many earlier models from DeepSeek, Kimi, MiniMax, Qwen, Zhipu and others read as Hopper-class (132 SMs), which the author says cannot distinguish H800, H100 or H200 chips. Some models, including Meituan's LongCat-2.0 and DeepSeek-V4.1-Flash, showed no GPU signature.
The author also addresses whether AMD hardware could produce a similar pattern. For DeepSeek-V4-Flash, he says 363 of 377 differing floating-point values match Nvidia and none match AMD's MI300X, and he calls AMD "very unlikely" for the other five. He estimates roughly 12,000 Blackwell GPUs across five training runs, with the largest clusters costing about $200-300 million, but says those figures are uncertain and were revised downward.
The report is explicit about its limits. The fingerprint shows the hardware used to initialize a model, which "strongly suggests" but does not prove training on the same hardware. It does not establish where the chips were, how access was obtained, or any export-control violation. The method only works on models with published weights and a hardware-dependent PyTorch pattern, and some results come from a single analysis pass that has not been independently re-run. If the technique holds up, open-weight releases could become a source of evidence in debates over chip export enforcement.