Weight forensics suggest DeepSeek, Kimi, MiniMax, Ant models were trained on banned Blackwell GPUs
SumitGup · x · 2026-11-13
Will Depue reports forensic evidence that DeepSeek, Moonshot AI, MiniMax, and Ant Group trained their newest models on NVIDIA Blackwell chips that are barred from export to China. The trick: model weights are initialized by the GPU using a random seed, and different GPU architectures produce distinct arrangements even with the same seed — a hardware fingerprint whose traces survive training. He searched over 6 trillion combinations, matching DeepSeek-V4-Flash/V4-Pro, Kimi K3/K3.1-Pro, MiniMax-M3, and Ling-3.1-flash to Blackwell (older models matched Hopper). He verified on real hardware, bit-exactly restoring DeepSeek-V4-Flash's initial weights on a B300 using the recovered seed (953049583). Unverified officially.
More from Infra
- 'Maybe the Great Filter Is Permitting a 10GW Data Center' Quip Resonates — Dr_Singularity · 2026-11-13
- Fab capacity jumps to 1.117M 8-inch-equivalent wafers/month, 78.9% on 12-inch — teortaxesTex · 2026-11-13
- Microsoft and Claude go down simultaneously, taking apps offline with them — ThePeterMick · 2026-11-13
- llama.cpp ships v1.0.0, the landmark first stable release for local LLM inference — TheApadayo · 2026-11-13
- Singapore files 41 charges against man in alleged Nvidia AI chip smuggling scheme to Malaysia — Polymarket · 2026-11-13
- Meme: All modern digital infrastructure rests on one tiny Model Hub pillar — pcuenq · 2026-11-13