Thirty-two lines of PyTorch decide where every starting weight in a large model lands, and one of those lines asks the GPU how many processor cores it has. That line is why a report published on 12 November can claim to read a model’s training hardware out of weights the labs themselves put online.
The report is checkpoint forensics, by Will DePue at Unemployment Labs. Its finding: six models from DeepSeek, Moonshot AI, MiniMax and Ant Group, all released since April 2026, carry initialisation patterns matching NVIDIA Blackwell chips, while earlier models from the same four labs match the older Hopper generation. The Blackwell parts named have required a US export licence for China since before they shipped, and the report states that none have been reported.
The statistics are not the weak point. There is a weaker link, and the report identifies it before any critic could.
How the fingerprint works
Before training can start, a model’s billions of weights need initial values. PyTorch generates them with a pseudorandom algorithm called Philox, spread across thousands of parallel GPU threads, and each thread writes into positions spaced by the total thread count. Change the number of threads and the same numbers land in different places.
The thread count comes from the chip. In the PyTorch file the report cites, the relevant line reads blocks_per_sm = maxThreadsPerMultiProcessor / block_size, with block_size_bound fixed at 256 just above it, and the grid capped at multiProcessorCount * blocks_per_sm. At 2,048 maximum threads per core that is eight blocks of 256, so total threads equal cores times 2,048.
I ran that multiplication rather than taking the report’s word for it. Hopper’s 132 cores give 270,336 threads; the B200 and B300’s 148 give 303,104; the GB200 and GB300’s 152 give 311,296. All three match the figures the report states exactly. The mechanism is real and a reader can check it in an afternoon.
What survives training is the harder part. Weights reserved for tokens never seen during training can keep their starting values almost untouched, and the report calls a match on these an exact copy. Trained weights keep only a faint statistical trace, reported for DeepSeek-V4-Flash at r = 0.066. Against roughly six trillion seed-and-layout combinations searched, the report sets about eight sigma as the bar a chance match could reach, and reports results from 49 to 1,385 sigma.
The chain has seven links, not one
“Model X was trained on Blackwell” is not one claim. Reading the report closely, it decomposes into seven, and they are closed to very different degrees:
- A seed and thread count that reproduce the published weights were recovered.
- The same combination predicts weights held out of the search.
- The AMD alternative is excluded. An MI300X normally uses 304 units but can be configured to 152 or 148, and the report confirmed on physical AMD silicon that it then reproduces the same layouts.
- The match is reproduced on physical Blackwell silicon.
- The initialisation is an independent event, not a seed inherited from an earlier model.
- Blackwell performed the training, not merely the initialisation.
- The chips were in China, or an export rule was broken.
The ledger
Counting the six models against those links, using the report’s own appendix:
- DeepSeek-V4-Flash, 148 cores, exact copy at 1,385 sigma. AMD excluded empirically: of 377 values where the two vendors’ arithmetic differs, the published weight matched NVIDIA in 363 and AMD in none. Reproduced on a physical B300, where unmodified PyTorch recovered 461 of 469 selected rows above r = 0.99, including 445 bit for bit.
- Kimi K3, 152 cores, exact copy at 305 sigma. Reproduced on a rented GB200, all 13 preserved rows, using Moonshot’s seed 20256586.
- MiniMax-M3, 148 cores, exact copy at 323 sigma.
- Ling-3.1-flash, 152 cores, untrained rows at r = 0.98, 244 sigma.
- DeepSeek-V4-Pro, 152 cores, partial survival at 49 sigma, the weakest of the six.
- Kimi K3.1-Pro, 152 cores, exact copy at 305 sigma, but starting from the same seed as K3 and, in the report’s words, it “may be a checkpoint of K3 with further post-training.”
So links 1 and 2 are closed for all six. Link 5 is closed for five. Link 4 is closed for two. Link 3, the AMD exclusion, is closed for exactly one, and the report says plainly that “the same test could not settle the other five models.” Links 6 and 7 are closed for none, which the report states rather than implies: the method identifies “the hardware that generated the starting weights, not every chip that later helped train the model,” and “by itself, it does not establish that chips were smuggled into China or that an export-control violation occurred.”
One model closes every link the method is capable of closing. The report’s own conclusion ranks it the same way, calling the evidence conclusive for DeepSeek-V4-Flash and strong for the others. That ordering is the first thing a headline count of six loses.
Where the report is harder on itself than a critic would be
Three places, and they are the reason the rest of it reads as credible.
It declines to count a seventh model that would have helped. Kimi K2.7-Code contains a part matching the 148-core layout, but the report finds the reserved rows are roughly 80% inherited Hopper values and 20% a new draw, identical in every row, which “points to a merged checkpoint rather than mixed hardware.” It is excluded from the six.
It publishes its failures. No signature was recovered from DeepSeek-V4.1-Flash “after an exhaustive search” — the newest model from the lab the report is most confident about. Baidu’s ERNIE-5.5 yielded a seed, 2760, but no hardware, because PaddlePaddle’s generator carries no chip layout. Meituan’s LongCat-2.0 gave no result.
It downgrades its own validation. The method agrees with all 24 cases where a lab has publicly named its hardware, which sounds like a 24-for-24 accuracy figure. The report refuses that reading: eleven are step-zero checkpoints and related models are not independent, so “this is a consistency check rather than a measured accuracy rate.” And it attaches to its own export-status column the sentence “They are not a finding that any law was broken.”
A caveat volunteered before anyone asks for it is worth more than the same caveat extracted later.
What I could check without the weights
Two corroborating artifacts the report cites are public, and both verify.
Ant Group’s cuLA repository does contain a benchmark file whose header reads GPU: NVIDIA GB300 | CUDA: 12.9 | PyTorch: 2.9.1+cu129, auto-generated on 2026-03-31, as described. A DeepSeek engineer did write, in DeepGEMM pull request 112 on 12 June 2025, “as we don’t have SM100 GPUs, so let’s make it a draft and wait for community feedback” — SM100 being Blackwell. Neither is evidence about the six models. Both are evidence that the report quotes its sources accurately where a reader can follow.
What I did not verify is the forensic result itself. That needs the weights, a Blackwell GPU and a Hopper GPU, and I have none of the three. The reproduction scripts and row fixtures are, per the report, “available to reviewers on request” rather than published, so nobody can currently re-run this from the page alone. The physical reproductions are the author’s measurements, reported by the author.
What weights cannot show
The forensics stop at the arrangement of numbers. A lab could generate weights on one cluster and train on another; the report says so. A mixed cluster of NVIDIA and Huawei parts would leave only the NVIDIA trace. Blackwell systems can be rented outside China, which makes access and location separate questions that weights cannot separate.
Those limits are not hedging. They are the difference between a forensic finding and a legal one, and the report keeps them apart on every page. The open question is not whether the fingerprints are there. It is link 3 on five of the six models, and the only published route to closing it runs through hardware the author has access to and the reader does not.