Researcher claims weight fingerprints prove 4 Chinese labs trained on banned Blackwell chips
_lewtun · x · 2026-11-13
Will Depue claims "conclusive forensic evidence" that DeepSeek, Moonshot, MiniMax and Ant Group trained their newest models on NVIDIA Blackwell chips barred from export to China. Method: initial weight layout patterns depend on the exact GPU and seed, leaving a hardware "fingerprint" that survives training in published weights. Commenters add that frontier open-source models are typically trained on only 2,000–3,000 chips, a vast scale gap vs. closed frontier. Unverified as of yet.
More from Infra
- SMIC hits record $3.19B Q3 revenue at 97.3% utilization, but capex lags plan — zephyr_z9 · 2026-11-13
- General Reasoning: $750K to prove out the West's first open multi-agent RL recipe for LLMs — rosstaylor90 · 2026-11-13
- Unsloth UD_IQ1_M Quant Saves 6GB but Breaks Tool Calls in Real Use — SomeoneInHisHouse · 2026-11-13
- One DGX Spark runs ripple-65b preview at 10.8 tok/s with coherent long reasoning — DZhang50 · 2026-11-13
- Qwen3.8-Flash-Next: 180B total / 6B active, runs on consumer GPUs via Strata offloading — drdanielbender · 2026-11-13
- Plotly's DataAnalyticsBench: three models that fit in 64GB score a B on 43 analytics tasks — DZhang50 · 2026-11-13