The GPU Leaves a Fingerprint: Checkpoint Forensics Ties Six Chinese Open-Weight Models to Blackwell

On 12 November 2026, Will DePue, a former OpenAI researcher who worked on Sora and on post-training GPT-4o and o3, posted a long thread on X and a report through his nonprofit, Unemployment Labs. His claim is that six open-weight models from four Chinese labs were initialized on Nvidia Blackwell hardware, which the US bars from export to China without a licence. The six models are DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ant Group’s Ling-3.1-flash. TechCrunch and The Decoder covered the report the next day. The verification code went up on GitHub at macrogrants/checkpoint-forensics on 13 November.

The political story is obvious. It arrives in the same week that Jensen Huang left Nvidia amid reporting about his ties to an executive at Megaspeed, a company implicated in chip smuggling. For developers, though, the more interesting part is the mechanism. It shows that a published checkpoint can carry metadata nobody intended to publish.

How a GPU leaves its signature in the weights

Before training starts, every weight needs an initial value. In PyTorch on Nvidia hardware, those values come from cuRAND’s Philox generator, which is deterministic: the same seed gives the same stream of numbers. To fill billions of values quickly, the work is spread across GPU threads. According to the report, a large enough initialization launches up to 2,048 threads per streaming multiprocessor (SM). Each thread draws its own subsequence and writes it into strided positions. The total thread count, and so the stride, depends on how many SMs the chip has:

  • Hopper (H100 SXM, H200, H800): 132 SMs, 270,336 threads
  • Blackwell B200/B300: 148 SMs, 303,104 threads
  • Grace Blackwell GB200/GB300: 152 SMs, 311,296 threads

With the same seed, a different SM count puts the same random numbers in different places. A random tensor therefore encodes the shape of the machine that generated it. This fits PyTorch’s own warning that random results are not guaranteed to be reproducible across platforms. Here that “non-reproducibility” becomes a signal.

The obvious objection is that training overwrites the starting values. DePue’s answer comes in two parts. First, some rows are never updated at all. Embedding rows for vocabulary tokens that never appear in the training data keep their initial values almost exactly. Second, ordinary trained weights keep a faint correlation with where they started. In 42 layers of DeepSeek-V4-Flash, the Blackwell layout predicts the final weights with r = 0.066, which is weak per weight but reaches 57 sigma across about 745,000 weights per layer. Under the Hopper layout, the correlation disappears.

The search covered roughly six trillion combinations of seed and layout. To guard against finding something just by trying enough options, each candidate is chosen on one set of weights and then tested on held-out weights it was not chosen on. The report puts the best expected accidental match at about 8 sigma. The real matches range from 49 to 1,385 sigma, and the weakest single test reported is 20 sigma.

The hardware check

The most persuasive part is not the statistics but the replication. With unmodified PyTorch and DeepSeek’s recovered seed (953049583, which DeepSeek reused from V3 to V4), a B300 reproduced 461 of 469 selected rows of DeepSeek-V4-Flash with correlation above 0.99, and 445 of them were bit-for-bit identical. An H100 reproduced all 57,344 preserved values of DeepSeek-V3 bit for bit, but none of V4-Flash. A rented GB200 reproduced all 13 preserved rows of Kimi K3 from Moonshot’s seed (20256586). Changing either seed by one broke the match. The report says the check runs on a laptop in seconds.

For calibration, the method matched all 24 models whose developers had publicly named their hardware. Llama 3.1 reads as 132 SMs (H100), Pythia reads as 108 (A100), and Nvidia’s own Nemotron-3 Ultra, partly trained on GB200, reads as 152. The report itself cautions that related models are not independent, so this is a consistency check rather than an accuracy rate.

The timeline in the report is also striking. DeepSeek reads as Hopper through V3 and as Blackwell from V4 (April 2026). Moonshot switches at K3 (July 2026), MiniMax at M3 (June 2026), and Ant between Ling-3.0-flash (August) and Ling-3.1-flash (October). The newest models from Alibaba’s Qwen, Zhipu, Tencent, Xiaomi and StepFun still read as Hopper. Meituan’s LongCat-2.0 carries no Nvidia fingerprint, which is consistent with its statement that it trained on “AI ASIC superpods”.

What it does not prove

The report is more careful than much of the coverage, and its caveats deserve attention:

  • Initialization is not training. The fingerprint shows which hardware generated the starting weights. A lab could initialize on one cluster and train on another. In the report’s words, the evidence “supports, without proving” that Blackwell did the training.
  • Location and legality are unknown. Blackwell capacity can be rented in Malaysia or Thailand. The evidence shows access, not smuggling, and not necessarily a legal violation. That depends on rules about who may use the chips, not only who owns them.
  • AMD is an alternative. An MI300X limited to 148 or 152 compute units reproduces the same layouts. DePue ruled this out for V4-Flash using a vendor-specific rounding difference: 363 of 377 values match Nvidia and none match AMD. For the other five models, the AMD explanation is only judged “very unlikely”.
  • The cost estimates are soft. The first version estimated 2,000–3,000 GPUs per large run. A same-day revision raised that to about 4,500 for Kimi K3 and 3,000 for DeepSeek-V4-Pro, or about 12,000 Blackwell GPUs and over half a billion dollars across five runs. These figures rest on assumed utilization, run length and token counts that the weights cannot confirm.
  • The verification is not independent yet. DePue says senior researchers at OpenAI and Anthropic confirmed the result, but none have been named. The full paper has not been published yet. No response from DeepSeek, Moonshot, MiniMax or Ant appears in the coverage reviewed here.

Why it matters

There are three takeaways for anyone who ships or uses open models.

First, open weights leak provenance. A checkpoint tells you not just what a model computes but how and on what it was created. That includes seed reuse across generations, merged checkpoints, and here hardware class. Kimi K2.7-Code, for example, shows a blend of about 80% Hopper-era rows and 20% newly drawn Blackwell-layout rows, a sign of a merged checkpoint. Expect “checkpoint forensics” to become a field of its own, and DePue says more findings are coming.

Second, this signal is easy to remove. Initializing on the CPU, using a different generator, or shuffling the tensor would all erase it. DePue acknowledges this and says he weighed that before publishing. Policymakers should treat it as a one-time look at the past, not a lasting verification tool.

Third, it adds evidence to the debate over whether export controls on chips can work when compute can be rented across borders. At least four leading Chinese labs moved to the newest Nvidia generation within months of each other, despite the ban. Controls that cover chip location but not chip use look leaky. The more durable route is the kind of compute governance that does not depend on labs forgetting to hide their tracks.

Sources