The Pulse

Researcher says six Chinese AI models carry Blackwell chip traces

Will DePue says patterns in six published AI models match layouts associated with Nvidia Blackwell GPUs. His analysis concerns the creation of starting weights and does not establish where the chips were located or whether export controls were violated.

AI.info Team · · fact-checked · against 3 sources ↓ · Published 13 Nov 2026

Six Chinese AI models contain numerical patterns matching Nvidia Blackwell GPUs used to create their starting weights, researcher Will DePue says. The models come from DeepSeek, Moonshot AI, MiniMax and Ant Group, according to his checkpoint-forensics report, as described by The Agent Times.

DePue's report was published on November 12. The finding concerns the hardware used to initialize model weights, not a direct identification of the machines used for the rest of training.

Which models matched

The analysis points to Blackwell B200 or B300 layouts for DeepSeek-V4-Flash and MiniMax-M3, The Agent Times reports. The other matches in the report are DeepSeek-V4-Pro, Moonshot's Kimi K3 and Kimi K3.1-Pro, and Ant Group's Ling-3.1-flash, which are assigned to the GB200 or GB300 layout.

The report also compares those results with earlier models from the same companies. DePue assigns DeepSeek-V3, Kimi K2 and K2.5, MiniMax-M2.7 and Ling-3.0-flash to Nvidia's older Hopper generation, according to The Agent Times.

DePue's code repository provides a way to replay checks for five models from the four companies, including the older DeepSeek-V3 comparison. It uses rows copied from published model weights and checks them against recorded hashes.

Reading the starting weights

Weights are the numerical values a model adjusts during training. Before training, software creates their initial values using a starting number known as a seed.

The method examines embedding rows associated with tokens that never appeared in training. Those rows receive no training gradient, so they can retain their initial values, the repository says.

The arrangement of those values depends on the number of streaming multiprocessors, or SMs, in the GPU used to create them. The repository's replay code rebuilds the rows using candidate SM counts and looks for a match with the published weights.

For DeepSeek-V4-Flash, the repository reports a match for 461 of 469 rows using the 148-SM layout associated with B200 and B300 chips. It says the remaining eight rows were partly trained. For Kimi K3, the repository reports that 13 of 13 rows matched the 152-SM layout associated with GB200 and GB300.

Hardware checks and limits

The repository also describes checks on physical GPUs using PyTorch. It reports matches on a B300 for DeepSeek-V4-Flash and MiniMax-M3, and on a GB200 for Kimi K3. A control using a different seed matched nothing on the tested GPUs.

A separate public repository reports reproducing the proposed layouts for all six models and DeepSeek-V3 using a central processing unit, JetsonsAge reports. Its checks do not include physical hardware tests, the AMD comparison or weaker traces in trained layers.

A specially configured AMD MI300X can also produce the relevant layouts, according to the report as described by JetsonsAge. DePue's additional numerical comparison favors Nvidia for DeepSeek-V4-Flash, while the AMD alternative remains open for the other five models.

The results do not establish where the chips were located, how access was obtained or whether export controls were violated, The Agent Times reports. The code repository likewise states that its check identifies the GPU that created starting weights, not where later training happened.