
Published AI model weights from DeepSeek, Moonshot AI, MiniMax and Ant Group contain numerical patterns consistent with NVIDIA’s restricted Blackwell GPUs, according to a forensic investigation published by former OpenAI researcher Will DePue on November 12.
The investigation examined six open-weight models released by the four Chinese AI developers since April 2026. It identified configurations matching NVIDIA’s B200, B300, GB200 and GB300 processors, hardware subject to US export licensing requirements for shipments to China.
The identified models include DeepSeek-V4-Flash and DeepSeek-V4-Pro, Moonshot AI’s Kimi K3 and Kimi K3.1-Pro, MiniMax-M3 and Ant Group’s Ling-3.1-flash. DePue published the findings through Unemployment Labs, drawing evidence from the numerical values preserved in the companies’ publicly released model weights rather than leaked documents or insider accounts.
DeepSeek-V4-Flash and MiniMax-M3 matched the 148-streaming-multiprocessor configuration associated with NVIDIA’s B200 and B300 GPUs. DeepSeek-V4-Pro, both Moonshot models and Ling-3.1-flash matched the 152-unit configuration associated with GB200 and GB300 systems. The method identifies compatible hardware configurations but cannot distinguish between individual GPU models within each group.
The technique examines how artificial intelligence models receive their initial numerical weights before training. PyTorch uses a seeded random-number generator called Philox to distribute values across GPU processing threads. Different GPU configurations produce different arrangements, even when the same random seed is used. Some values remain unchanged during training, allowing researchers to compare published weights with reconstructed initial values and identify the original hardware configuration.
DePue searched approximately six trillion combinations of random seeds and hardware layouts, then tested successful matches against weights excluded from the initial search. The reported statistical evidence ranged from 49 to 1,385 standard deviations across the six models.
Physical hardware tests also reproduced the patterns in two cases. Running unmodified PyTorch on an NVIDIA B300 reproduced 461 of 469 selected rows from DeepSeek-V4-Flash with correlations above 0.99. A rented GB200 reproduced all 13 selected rows from Kimi K3. The H100 tests did not reproduce the Blackwell-associated patterns.
Researcher Bojie Li independently implemented the verification method without using DePue’s code and examined 265 publicly available open-weight models. His findings matched DePue’s hardware classifications across 65 models examined by both researchers, with no disagreements. Both also found no identifiable GPU fingerprint in nine models.
The investigation does not establish where the processors were located, who owned them or how the developers obtained access. It also cannot confirm that the same hardware performed every stage of model training. DePue ruled out specially configured AMD hardware as an alternative for DeepSeek-V4-Flash but considered it very unlikely, rather than conclusively excluded, for the other five models.
Kimi K3.1-Pro also shares its initialisation seed with Kimi K3, suggesting that it may be a further post-trained version of the earlier model rather than an independent training run.
Discover more from Aree Blog
Subscribe now to keep reading and get access to the full archive.


