Report links 6 Chinese AI models to Nvidia Blackwell GPUs
Researcher Will DePue says initial weight patterns in models from DeepSeek, Moonshot AI, MiniMax and Ant Group match export-restricted Blackwell chips.
Researcher Will DePue of the nonprofit Unemployment Labs published a report on November 12, 2026, presenting forensic evidence that six models from four Chinese AI companies were initialized on export-restricted Nvidia Blackwell GPUs, according to aitimes.com.
The six models analyzed are DeepSeek's V4-Flash and V4-Pro, Moonshot AI's Kimi K3 and K3.1 Pro, MiniMax's M3, and Ant Group's Ling 3.1 Flash. DeepSeek V4-Flash and MiniMax M3 matched the compute unit configuration of Blackwell B200 and B300, while the other four matched GB200 and GB300 patterns, with statistical evidence ranging from 49 to 1385 sigma.
In reproduction experiments, 445 rows of DeepSeek-V4-Flash initial weights matched bit-for-bit on a Blackwell B300, and all 13 rows of Kimi K3 were reproduced on a GB200. The same DeepSeek-V4-Flash pattern did not appear on Hopper H100. The method was validated on 24 existing AI models whose GPUs were disclosed, and all matched the published information.
The analysis used only public model weights, not insider tips or leaked data. The researchers also found a record in Ant Group's published program code of a performance measurement on a Blackwell GB300 dated March 31, 2026.
The report does not prove the Blackwell chips were used for full training, nor does it establish where the GPUs were installed, who owned them, or how they were obtained.
Shift from Hopper to Blackwell
DeepSeek's V3, released in 2024, showed Hopper-family traces, but V4, released in April 2026, showed Blackwell patterns. Moonshot AI's Kimi K2.5 from January 2026 was Hopper-based, while K3 from July 2026 showed Blackwell patterns. US government officials had previously raised suspicions that DeepSeek and Moonshot AI used Blackwell chips.
Quick answers
Which models were analyzed in the report?
DeepSeek's V4-Flash and V4-Pro, Moonshot AI's Kimi K3 and K3.1 Pro, MiniMax's M3, and Ant Group's Ling 3.1 Flash.
Does the report prove the chips were used for training?
No. The report does not prove the Blackwell chips were used for full training, nor does it establish where the GPUs were installed, who owned them, or how they were obtained.
What data did the researchers use?
Only public model weights, not insider tips or leaked data.