Skip to content- Sources: primary, discussion
- Summary: The report regenerates initial weights from cuRAND Philox seeds and tests which thread counts reproduce traces that survive training. The thread count depends on a GPU's SM count: 132 for Hopper, 148 for B200 and B300, and 152 for GB200 and GB300. DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash match Blackwell, while earlier models from the same labs match Hopper. The method agreed with all 24 models whose hardware was publicly disclosed. It shows only where the weights were initialized, not where the chips were or whether controls were broken.
- Why it matters: The released weights of an open model can reveal the hardware that initialized them, which matters for provenance checks and for anyone publishing weights.
- Background: PyTorch fills new GPU tensors with Philox, a counter-based random generator from cuRAND, split across up to 2,048 threads per streaming multiprocessor. The same seed on a chip with a different SM count puts the same numbers in different positions. Weights that training barely touches, such as unused vocabulary rows, keep those positions.
- Comments: The Hacker News thread drew 280 points and 82 comments.