Analysis finds Blackwell fingerprints in six open-weight models from four Chinese labs

Six open-weight models released since April 2026 by DeepSeek, Moonshot AI, MiniMax and Ant Group contain numerical patterns that match NVIDIA Blackwell GPUs, according to a forensic analysis by Will DePue of Unemployment Labs published November 12, 2026. Earlier models from the same labs match NVIDIA's older Hopper generation. Blackwell chips require a U.S. license for export to China, and the report notes that none have been reported.

What the analysis found

The six models read as Blackwell-initialized:

• DeepSeek-V4-Flash (April 2026): Blackwell B200/B300
• DeepSeek-V4-Pro (April 2026): Blackwell GB200/GB300
• Kimi K3 (July 2026): Blackwell GB200/GB300
• Kimi K3.1-Pro (November 2026): Blackwell GB200/GB300
• MiniMax-M3 (June 2026): Blackwell B200/B300
• Ling-3.1-flash (October 2026): Blackwell GB200/GB300

The B200 and B300 have 148 streaming multiprocessors (SMs); the GB200 and GB300 have 152. Kimi K3.1-Pro starts from the same seed as K3 and may be a checkpoint of K3 with further post-training.

The prior generation reads as Hopper: DeepSeek-V3 (December 2024), Kimi K2 (July 2025), Kimi K2.5 (January 2026), MiniMax-M2.7 (April 2026) and Ling-3.0-flash (August 2026). A Hopper fingerprint cannot distinguish an H800 from an H100 SXM or H200. DeepSeek and Moonshot disclosed H800 use for their Hopper-era models; that is a lab disclosure, not a finding of the method.

Kimi K2.7-Code, posted six weeks before K3, is built on Hopper-era weights but contains a small part that matches the B200 or B300.

Not every Chinese lab has moved. The newest models the author could read from Alibaba (Qwen), Zhipu (GLM), Tencent, Xiaomi and StepFun all read as Hopper-trained. Meituan says LongCat-2.0 was trained entirely on "AI ASIC superpods," and the report found no NVIDIA fingerprint in its released weights.

How the method works

Before training, model weights need initial values. In PyTorch, these come from the Philox pseudorandom number generator, run through NVIDIA's cuRAND library. The work is split across GPU threads. On the chips examined, a sufficiently large initialization uses the maximum of 2,048 threads per SM, so a chip with a different SM count assigns the same job differently. The report recovers traces of the initial values and searches for the hardware configurations and seeds that reproduce them.

For DeepSeek-V4-Flash and Kimi K3, the match was also confirmed on physical Blackwell chips.

Validation and confidence

For 24 models where the developer has publicly stated the hardware used, the method agreed with the announcement in every case. NVIDIA's Nemotron-3 Ultra, whose technical report names the GB200 for one phase of training, reads as 152 SMs, the same as Kimi K3.

Across roughly 6 trillion seed-and-layout combinations searched, the strongest chance match would be expected to reach about 8 sigma. The strongest evidence for each of the six models ranges from 49 to 1,385 sigma, and the weakest individual test reported reaches 20 sigma.

The fingerprint identifies how the weights were created. It supports, without proving, the inference that Blackwell also performed the training. Specially configured AMD hardware, also export-restricted, is the only alternative identified. The report excludes it for DeepSeek-V4-Flash and considers it very unlikely for the other five.

Scale and cost estimates

From model size and training data, the report estimates each run needed roughly 500 to 3,000 Blackwell GPUs for two to three months, assuming each model was trained on the hardware that initialized it. Applied to DeepSeek-V3, the same arithmetic gives 2,000 to 2,500 chips; DeepSeek disclosed 2,048.

At roughly $70,000 per GPU, a cost implied by a Wall Street Journal report that about 36,000 Blackwell chips being installed in Malaysia for ByteDance's use would cost more than $2.5 billion, the largest of these clusters would cost $150 million to $200 million. The report puts the five together at more than half a billion dollars, and says total spending on export-controlled GPUs is likely in the billions. These estimates cover final training runs only.

What this means

The report adds technical evidence to earlier allegations. U.S. officials and press reports had said DeepSeek and Moonshot AI were using Blackwell chips, without publishing technical evidence. Here the evidence comes from the models' published weights rather than leaks or insider accounts.

The limits are stated by the author. The findings show access to advanced hardware, but do not establish where the chips were located, how access was obtained, or whether export controls were violated. The Hopper readings also cannot separate permitted chips from others of the same generation. The result that matters most for practitioners is narrower: weights alone can carry a verifiable record of the hardware that initialized them.