Unemployment Labs published a report on November 12, 2026, that attributes the starting weights of six publicly released models to patterns matching NVIDIA Blackwell GPUs. The work gives readers a way to assess claims about hardware access using information that survived in the models themselves.
What the report found
Will DePue examined model weights from DeepSeek, Moonshot AI, MiniMax and Ant Group. The report identifies a Blackwell pattern in DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash. It reports Hopper patterns for some earlier models from the same developers, including DeepSeek-V3 and Kimi K2.5.
The report says the work used the model developers' published weights. DePue looked for traces of the numbers used when a model was initialized, then checked whether particular hardware layouts produced the same pattern on weights that were left out of the search. He reports confirming a match on physical Blackwell systems for DeepSeek-V4-Flash and Kimi K3.
What a starting signature does not show
The chip used to initialize a model need not be the only chip used to train it. The report describes that distinction expressly. It also says the observations cannot place the equipment in a particular country or show how the developers obtained access. Those are different questions that would require evidence beyond model weights.
Where the test applies
The analysis needs weights that researchers can inspect and a setup that leaves a recognizable signature. A model for which no match was found may still have used a GPU. The report says it could not identify a GPU signature for some models, so readers should treat an unresolved case as an unknown rather than a finding about the hardware.
