ResearchNews
Model weights point to Blackwell GPUs at Chinese labs
Former OpenAI engineer Will DePue showed from public weights that six models from four Chinese labs were initialized on export-controlled NVIDIA Blackwell chips. The method reads traces left by the initial random values. Independent teams have already partly reproduced it.
· 4 min read

What the report claims
In a report published November 12, Will DePue says six models released since April 2026 by DeepSeek, Moonshot AI, MiniMax and Ant Group carry patterns matching Blackwell chips (Unemployment Labs). Earlier models from the same labs match Hopper.
- B200/B300 (148 SMs): DeepSeek-V4-Flash, MiniMax-M3.
- GB200/GB300 (152 SMs): DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, Ling-3.1-flash.
- Still Hopper: the newest readable models from Alibaba (Qwen), Zhipu (GLM), Tencent, Xiaomi and StepFun.
Kimi K3.1-Pro starts from the same seed as K3, so it is likely a further-trained checkpoint of it (Kimi K3.1 Pro: Moonshot's open-weight model nears Opus).
How weights reveal the chip
PyTorch fills initial weights with the Philox random generator via cuRAND, spread across many parallel GPU threads. Each SM runs 2,048 threads, so the thread count differs by chip generation: 270,336 on Hopper, 303,104 on B200/B300 and 311,296 on GB200/GB300. The same seed with a different thread count puts the same values in different positions.
Weights that training never touches, such as embedding rows for tokens never seen, keep their starting values almost exactly. DePue searched about 6 trillion seed and layout combinations. The expected best chance match is about 8 sigma; the strongest evidence for the six models ranges from 49 to 1,385 sigma. Trained weights keep a faint trace too: across 42 layers of DeepSeek-V4-Flash, r = 0.066 under the Blackwell layout and essentially zero under Hopper.
The method matched all 24 cases where a developer had disclosed the hardware. NVIDIA's own Nemotron-3 Ultra reads as 152 SMs, the same as Kimi K3.
Who checked it
- On a rented GB200, unmodified PyTorch reproduced all 13 preserved rows of Kimi K3. Changing the seed by one reproduced none.
- The Institute for Progress reproduced the DeepSeek-V3, DeepSeek-V4-Flash and MiniMax-M3 results on real GPUs. It lacked a 152-SM GPU to fully test Kimi K3 (
- [object Object]
- ).
- A separate replication recovered the DeepSeek-V3, V4-Flash and Kimi K3 layouts with correlations above 0.999 (
- [object Object]
- ).
- Bojie Li scanned 265 open models and reached the same GPU class for 65 models where both had a reading, with no contradictions (
- [object Object]
- ).
What it does not prove
The report is explicit: the fingerprint identifies the hardware that generated the starting weights. It does not show where the chips were, who owned them, or whether export rules were broken. Blackwell capacity can be rented outside China. A specially configured AMD MI300X could in theory produce the same layout; DePue excluded that for DeepSeek-V4-Flash and considers it very unlikely for the others. He calls only the DeepSeek-V4-Flash result conclusive.
He also estimates that the five Blackwell runs together needed roughly 12,000 GPUs, while saying he is not confident in that figure because run lengths and utilization are not public. The report lands in an already turbulent week for NVIDIA (Jensen Huang resigns; Colette Kress is NVIDIA's interim CEO).
Why it matters
From a security angle this is a new side channel: open weights reveal not only the model but also the hardware that created it. Anyone releasing open weights should assume that untouched embedding rows can expose the seed and GPU generation. For the export-control debate, it is the first public technical evidence anyone can verify, rather than a leak.
Key takeaways
- Starting weights of six Chinese models point to Blackwell GPUs; earlier ones to Hopper.
- The method reads the Philox/cuRAND thread layout from the weights, and the code is public.
- Independent teams reproduced the result for several models.
- It does not prove where the chips were or that any rule was broken.
Produced with the help of AI, reviewed by an editor. How we make our articles
Reviewed and edited by the Neural Daily desk. About


