Open weights leak which GPU trained them — six Chinese models read as NVIDIA Blackwell
Antique_Atmosphere97 · reddit · 2026-11-13
A forensics writeup exploits PyTorch's cuRAND/Philox weight initialization: the 2,048-threads-per-SM layout determines where each random number lands, and rows training never touches keep initial values, letting you recover the SM count of the initialization GPU. The method matches all 24 models with disclosed hardware (Llama 3.1 = 132 SMs/H100; Nemotron-3 Ultra = 152/GB200) and reproduces 445 rows of DeepSeek-V4-Flash bit-for-bit on a B300. Six recent Chinese models (DeepSeek-V4-Flash/Pro, Kimi K3/K3.1-Pro, MiniMax-M3, Ling-3.1-flash) all read as Blackwell (148/152 SMs) — the author notes this reveals initialization hardware only, not chip location or rule violations. Practical takeaway: publishing a checkpoint leaks your hardware unless you init on CPU (some Qwen3 did) or avoid this layout.
More from Infra
- Report: Nvidia planning frontier-class open models, sources claim — StewartalsopIII · 2026-11-13
- Celesto CI runs GitHub Actions on microVMs for ~90% less per minute — aniketmaurya · 2026-11-13
- China deploys 30,000+ workers to build massive AI data center hub in Inner Mongolia — Polymarket · 2026-11-13
- Lawyer runs DeepSeek Flash at near-NVFP4 quality on eight 2018 V100s: 524K context, 100+ tok/s, fully local — TumbleweedNew6515 · 2026-11-13
- "Singapore" reportedly accounts for 15-20% of NVIDIA revenue, and the jokes write themselves — JensHonack · 2026-11-13
- SGLang Summit agenda: LinkedIn serves ads model at 10K+ QPS under 80ms — DavidBennett__ · 2026-11-13