Will DePue published a report on the Unemployment Labs site on 12 November that relies solely on the models’ published weights. The report finds the Blackwell fingerprint in DeepSeek-V4-Flash and V4-Pro, Moonshot AI’s Kimi K3 and K3.1-Pro, MiniMax-M3 and Ant Group’s Ling-3.1-flash. Earlier models from the same companies carry a fingerprint that matches Nvidia’s older Hopper generation. DePue’s X profile lists earlier work at OpenAI, CellCog reported.

How a chip leaves a fingerprint

Before training, a generator fills the model weights with random numbers, which it produces reproducibly from a seed. PyTorch splits the work across the GPU’s threads. The number of threads depends on how many compute units, or SMs, the chip has: Hopper has 132, B200 and B300 chips have 148, and GB200 and GB300 chips have 152. The same seed therefore places the same numbers in different positions on different chips.

Some weights stay almost unchanged, such as the rows for words not used in training. Weights that change during training can also correlate weakly with their initial values, but the fingerprint does not survive every training run. DePue searched for the seed and thread count that reproduce these traces, then tested the combination on weights not used in the search.

Results and their limits

Under the report’s statistical model, a search of about 6 trillion combinations would by chance produce a match of about 8 sigma at best. The strongest results for the six models range from 49 to 1,385 sigma. On a B300 chip, DeepSeek’s seed reproduced 461 of 469 selected rows in V4-Flash with a correlation above 0.99. A rented GB200 reproduced all 13 preserved rows of Kimi K3. The method was right in all 24 cases where the developer has disclosed the model’s hardware. DePue regards this as a consistency check, not a measure of accuracy.

The fingerprint shows only which hardware generated the initial values. The report says it supports, but does not prove, the conclusion that the models were also trained on Blackwell chips. The weights do not reveal who owned the chips or where they were. Blackwell systems can be rented outside China. The report therefore does not show smuggling or any breach of export controls.

The only other hardware identified that can produce the same fingerprint is the export-controlled AMD MI300X, when its use is limited to 148 or 152 compute units. For V4-Flash, DePue ruled it out on the basis of small differences in the numbers. For the other five models, DePue considers it very unlikely.

Claims on X and scale estimates

DePue called the evidence indisputable in a thread on X, CellCog and The Agent Times reported. The report uses the word only for V4-Flash. CellCog also reported that the thread says researchers at OpenAI and Anthropic confirmed the result. The report does not name them. The code is not public; the report says reviewers can obtain the scripts on request.

The report estimates that each of the five training runs needed on the order of 1,500–3,500 Blackwell GPUs for one and a half to three months. It describes the calculation as uncertain. The calculation assumes that training took place on the same hardware that generated the initial values. Unite.AI and CellCog cite lower figures from the same report; Unite.AI, for example, gives 500–3,000 GPUs per run.