Six Chinese Models Read as Blackwell-Trained in a Seed Fingerprint Report

Six Chinese language models released between April and October 2026 were initialised on Nvidia Blackwell GPUs, a hardware generation U.S. export rules bar from China, according to a report published Nov. 12 by Will DePue through Unemployment Labs. The claim rests on a trace the weights themselves carry, not on a leak.

The report names DeepSeek, Moonshot AI, MiniMax and Ant Group, and gives release dates for the switch: DeepSeek with V4 in April 2026, MiniMax with M3 in June, Moonshot with K3 in July and Ant with Ling-3.1-flash in October. It says earlier models from the same four labs read as Hopper, Nvidia's older generation, and that each lab switched with one release.

How a random seed names a chip

Model weights begin as random numbers drawn by PyTorch's Philox generator from a seed. The draw is split across GPU threads, and the thread count decides which value lands in which slot. A given seed therefore produces a different starting pattern on different hardware.

The report lists three layouts. Hopper parts (H100 SXM, H200, H800) run 132 streaming multiprocessors and 270,336 threads. Blackwell B200 and B300 run 148 and 303,104. GB200 and GB300 run 152 and 311,296.

DePue looks for surviving starting values in weights that training barely changed, then tests candidate seeds on held-out weights. As a check, an unmodified PyTorch run on an H100 reproduced all 57,344 preserved DeepSeek-V3 values bit for bit. A B300 reproduced 461 of 469 selected DeepSeek-V4-Flash rows, 445 of them bit for bit. A rented GB200 reproduced all 13 preserved Kimi K3 rows.

The report puts the chance of a coincidental match at about 8 sigma across roughly 6 trillion seed-and-layout combinations. The six Blackwell-linked models score 49 to 1,385 sigma, and the weakest single test reaches 20.

The six models and the chips they read as

The table is the report's own classification. DePue calls the evidence conclusive for DeepSeek-V4-Flash.

Model	Lab	Layout it reads as
DeepSeek-V4-Flash	DeepSeek	148 SMs (B200/B300)
MiniMax-M3	MiniMax	148 SMs (B200/B300)
DeepSeek-V4-Pro	DeepSeek	152 SMs (GB200/GB300)
Kimi K3 and K3.1-Pro	Moonshot AI	152 SMs, same seed
Ling-3.1-flash	Ant Group	152 SMs (GB200/GB300)

The newest Qwen, GLM, Tencent, Xiaomi and StepFun models read as Hopper. Meituan's LongCat-2.0 shows no Nvidia fingerprint, and its model card refers to "AI ASIC superpods", the report says.

What 12,000 GPUs would cost

The scale figures are the weakest part, and the author says so: "We are not confident in these estimates." They assume one and a half to three months per training run.

1. Kimi K33400 GPUs
2. DeepSeek-V4-Pro2700 GPUs
3. DeepSeek-V4-Flash1700 GPUs
4. MiniMax-M31400 GPUs
5. Ling-3.1-flash1400 GPUs

Source: Unemployment Labs, Table 2, accessed 2026-11-13

The five clusters total about 12,000 GPUs, in a range of 7,000 to 20,000. At roughly $70,000 a GPU, a figure the report derives from a Wall Street Journal item on ByteDance chips in Malaysia, the largest cluster comes to about $200 million and all five to more than half a billion dollars.

Those are the author's numbers, built on an assumed run length and an assumed price, and the report says as much in its own caveats. The denominator, GPU-months per run, is exactly the input a lab could dispute.

What the fingerprint does not show

The trace marks where the starting weights were generated, not where every step of training ran. Weights could be initialised on one system and trained on another, and a mixed cluster could hide older parts. The method cannot name an owner, a location or a route. DePue writes that the evidence "concerns access to the hardware" and does not by itself show smuggling or an export violation.

It also works only on published weights initialised with GPU PyTorch. Closed models, several CPU-initialised Qwen3 models and PaddlePaddle models such as ERNIE-5.5 cannot be checked. An AMD MI300X can be configured to mimic these layouts; for DeepSeek-V4-Flash, 363 of 377 disputed values matched Nvidia's output and none matched AMD's.

What was already alleged, and what would change this read

The report lands on top of an older argument, and the older argument had no method attached. On Feb. 23, 2026, a senior Trump administration official said DeepSeek's coming model was trained on Blackwell chips, likely clustered in Inner Mongolia, Reuters reported. The official did not say how the U.S. obtained the information. That claim had no public evidence behind it; this one has a method others can rerun.

Nvidia is in the middle of a leadership change, covered in our report on Jensen Huang's resignation, and the export question will follow its next chief executive. For buyers of accelerators the more immediate consequence is demand: inference-chip startups such as Fractile sell on the premise that supply is constrained.

What would change this read is a rerun by a party with no stake. A second lab reproducing the B300 match on DeepSeek-V4-Flash would settle the hardware question. Evidence that the weights were seeded on a development box and trained elsewhere would shrink the 12,000-GPU estimate toward zero, and the report's own caveat allows for it.