AI news, problem solutions & tool picks
DigestX

AI News

Six Chinese open-weight models carry the fingerprint of Nvidia Blackwell chips, a report says

· 2 min read

Photo via Pexels

Six Chinese open-weight models carry the numerical fingerprint of Nvidia Blackwell graphics processors, according to a report published on November 12 by Will DePue, a former OpenAI researcher who now runs the nonprofit Unemployment Labs. The models are DeepSeek-V4-Flash and DeepSeek-V4-Pro, Moonshot AI’s Kimi K3 and Kimi K3.1-Pro, MiniMax-M3 and Ant Group’s Ling-3.1-flash. Blackwell chips are barred from export to China under US rules.

How the fingerprint works

Before training begins, a model’s weights are initialised with a pseudorandom generator called Philox, and PyTorch spreads that work across thousands of threads on the GPU. Counting the values pinned in the published weights recovers the thread layout: 148 units for the B200 and B300 chips, 152 for the GB200 and GB300. DePue searched more than six trillion combinations of chip layout and seed number. On a real B300, he reproduced 445 of 469 selected weight rows of DeepSeek-V4-Flash bit for bit; on an H100, none matched. The same labs’ older models — DeepSeek-V3, Kimi K2.5 — read as Hopper, Nvidia’s previous generation. Alibaba’s Qwen, Zhipu’s GLM, Tencent, Xiaomi and StepFun still point to Hopper.

What it does not prove

The fingerprint identifies the chip that generated each model’s starting weights, before training began. DePue’s own report says the evidence “supports, without proving” that training ran on Blackwell. It does not establish where the chips were, how the labs got access to them, or whether any export rule was broken. DePue estimates the five Blackwell runs needed roughly 12,000 GPUs between them, and that the two largest clusters alone would cost $200–300 million at about $70,000 a chip — those are the report’s calculations, not direct measurements. The report calls the evidence “conclusive” only for the starting weights of DeepSeek-V4-Flash.

The method needs no leaks or insiders — it reads only the weights the labs themselves publish. That is what makes it hard to dismiss and hard to extend: it can say where a model began, and nothing about where the rest of its training happened.