OpenAI researcher claims weight forensics show Chinese labs trained on banned Blackwell GPUs
willdepue · x · 2026-11-13
OpenAI employee Will Depue (unverified, personal claim) says forensic analysis of published model weights shows DeepSeek, Moonshot, MiniMax, and Ant Group trained their newest models on export-banned NVIDIA Blackwell chips.
- Method: no insider info—weights themselves carry a hardware fingerprint, since initialization patterns depend on the exact GPU+seed combination and survive training.
- Claims: DeepSeek-V4-Flash/V4-Pro, Kimi K3/K3.1-Pro, MiniMax-M3, and Ling-3.1-flash used Blackwell B200/B300 or GB200/GB300; earlier versions ran on Hopper.
- Next: he says a GitHub repo and full paper with replication details are coming.
This is a one-sided allegation, independently unverified; NVIDIA and the labs have not responded.
More from Models
- Jacques Thibs: AI's 1.5-year capability jump came mostly from unhobbling — JacquesThibs · 2026-11-13
- Gemini 4 Hydrogen debuts at #4 on LLM-to-3D benchmark, third Google model in top 4 — DynamicWebPaige · 2026-11-13
- Power user calls out OpenAI's slow shipping: Anthropic released Opus 5.5 and Fable 5.2 — Angaisb_ · 2026-11-13
- airbench: Testing Local Models on Visual Desktop Challenges via Keyboard and Mouse Agents — dh7net · 2026-11-13
- vLLM hits 58x GB300 throughput on Rubin with DeepSeek-V4.1-Flash optimization — vllm_project · 2026-11-13
- vLLM hits 58x GB300 throughput on Rubin with DeepSeek-V4.1-Flash optimization — vllm_project · 2026-11-13