A newly published research report says that six open-weight models released this year by DeepSeek, Moonshot AI, MiniMax and Ant Group carry initialization traces of Nvidia Blackwell GPUs in their weights. The report was written by Will DePue, head of the nonprofit research group Unemployment Labs, whom TechCrunch describes as a former OpenAI researcher. Blackwell is among the chips the United States bars from export to China.
Where the fingerprint comes from
The method looks at the step before training begins: random initialization. When PyTorch generates initial weights on a GPU, it uses the Philox random number algorithm, and where each value lands depends on the number of streaming multiprocessors (SMs) and threads on the card. Each chip generation is configured differently: the B200/B300 has 148 SMs and 303,104 threads, the GB200/GB300 has 152 SMs and 311,296 threads, and the previous-generation Hopper has 132 SMs.
The author searched roughly 6 trillion “seed plus layout” combinations, then validated the result against weights that were not part of the search. Training overwrites most weights, but some rows always keep their initial values. According to the report, unmodified PyTorch on an H100 reproduces those rows in DeepSeek-V3, while on a B300 it reproduces them in DeepSeek-V4-Flash; swap the two and nothing matches. On a rented GB200, the author also reproduced the surviving rows of Kimi K3.
Six models, one dividing line
The six models listed in the report are DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ant Group’s Ling-3.1-flash. Their statistical significance ranges from 49 to 1,385 sigma, against a random baseline that the author puts at about 8 sigma.
Line up the release order of these companies and the split is fairly clean. Earlier models such as DeepSeek-V3, Kimi K2.5, MiniMax-M2.7 and Ling-3.0-flash all read as Hopper; the newer generation, released in succession from April 2026 onward, switches to a Blackwell fingerprint. Kimi K2.7-Code sits between the two with mixed results and was not counted.
Other companies look different. Models from Qwen, Zhipu AI’s GLM, Tencent, Xiaomi and StepFun still read as Hopper in the report, while Meituan’s LongCat-2.0 yielded no GPU fingerprint; Meituan has previously said it was trained on an “AI ASIC supernode.” Within the scope of the report, the Blackwell traces are concentrated in four companies and do not extend across China’s entire open-source scene.
The report also estimates the GPU scale of the final training run: a best estimate of about 4,500 GPUs for Kimi K3, about 3,000 for DeepSeek-V4-Pro and around 1,500 each for the rest, roughly 12,000 in total. At about $70,000 per GPU, the largest cluster would be worth between $200 million and $300 million. The author is explicit about being unsure of these figures, which depend on assumptions about compute utilization, the number of training tokens and training duration, and the numbers in the first version of the report were revised once on the day of publication.
The limits the report sets for itself
What this research can show is limited, and the author lists the caveats in detail. The fingerprint only records which kind of chip generated the initial weights; a model can easily be initialized on one system and trained on another. AMD’s MI300X could in theory mimic these layouts, and the author ran a numerical exclusion only for DeepSeek-V4-Flash, calling the other five “highly unlikely.”
More sensitive still, the method cannot tell who owns the chips or where they are. Blackwell servers can be rented in data centers outside China, and in the report’s own words, the evidence does not prove the chips were smuggled into China, nor that any violation of export controls took place. On the training stage, the report’s wording likewise leaves room:
“supports, without proving, the inference that Blackwell also performed the training”
The method only works on models with public weights; closed models cannot be tested. TechCrunch’s coverage includes no response from the Chinese companies or from Nvidia, and the US Department of Commerce has not commented on the study, so the claims cannot currently be verified with the parties involved.
In the same week the report appeared, the US Department of Commerce was reported to be finalizing rules that would restrict Chinese AI companies from remotely using advanced chips through overseas data centers. If those rules take effect, renting Blackwell abroad would be the first route hit, and this study offers a way to check that relies on no insider information, only public weights. When the next batch of open models releases its weights, someone is likely to run the same test again.
Sources: TechCrunch, CocoLoop, Unemployment Labs research report Checkpoint Forensics; the report text was checked for the list of six models, SM and thread counts, the sigma range and the basis of the GPU-scale estimates.