Loading page…
Will DePue’s analysis traces patterns left when model weights were first created. It suggests access to Blackwell hardware, but cannot establish where the chips were or how access was obtained.
Listen to this story
A Nov. 12 analysis by former OpenAI researcher Will DePue uses residual patterns in released model weights to identify two Nvidia Blackwell hardware configurations associated with six Chinese models. The matches offer evidence about the GPUs used to generate models’ starting values, but do not prove those GPUs trained the models or establish where the hardware was located. The work adds a possible way to scrutinize compute claims from model releases; it does not, by itself, show chip smuggling or an export-control violation.
DePue associates DeepSeek-V4-Flash and MiniMax-M3 with B200/B300 configurations, and DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash with GB200/GB300 configurations.
The two inferred configurations have 148 or 152 streaming multiprocessors; the method cannot identify a specific chip within either group.
DePue says his method matched disclosed hardware for 24 control models; a B300 reproduced V4-Flash’s untouched starting weights bit for bit, while a GB200 matched Kimi K3 at 0.999995 correlation.
Six Chinese AI models contain numerical traces that point to access to Nvidia’s restricted Blackwell chips, researcher Will DePue says. In an analysis published November 12, he links models from DeepSeek, Moonshot AI, MiniMax and Ant Group to the hardware through their publicly released weights—not insider accounts or shipping records.
DePue’s method identifies the hardware configuration used to create a model’s starting values. It supports, but does not prove, that the same hardware trained the model.
Weights are the billions of adjustable numbers that shape a model’s answers. Before training, software fills them with initial values generated from a starting number called a seed. The same seed and settings reproduce the same sequence.
PyTorch, the model-development framework, divides that work among parallel workers on a GPU. Different hardware configurations use different worker counts, placing the generated numbers in different positions. Recovering the arrangement requires both the right seed and the right layout.
Some weights for words or symbols never encountered during training remain almost untouched. Others retain a faint statistical relationship with their starting values. DePue searches these surviving traces, which his report cautions are not guaranteed to survive every training run.
The six models fall into two groups in DePue’s analysis, distinguished by their number of processing units, called streaming multiprocessors. The method cannot distinguish individual chip models within each group.
Earlier models from all four labs matched Nvidia’s older Hopper generation, according to the report. DePue also notes that Kimi K3.1-Pro uses the same seed as K3 and may be a version with further post-training—not a separate training run from scratch.
DePue tests recovered settings against weights withheld from the search. He says the method agreed with developers’ hardware disclosures for all 24 control models.
In his November 12 X post, he says a Blackwell B300 reproduced V4-Flash’s untouched starting weights bit for bit. A rented GB200 reproduced Kimi K3’s with a correlation of 0.999995.
He also claims verification by unnamed senior researchers at OpenAI and Anthropic. The Decoder reports that he published verification code on GitHub.
The report identifies specially configured AMD hardware as an alternative explanation, though DePue says he excluded it for V4-Flash and considers it very unlikely for the other five models.
Chinese companies can rent computing capacity outside China. A hardware fingerprint in released weights is not a shipping record: even a correct match cannot, by itself, establish physical smuggling or an export-law violation.
Loading discussion...
Join the conversation
Explain where you would draw the line between a useful lead and sufficient evidence.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.