claim C-0044 Unconfirmed
DePue report: six models from four labs were initialized on Blackwell-class hardware (148/152 SMs); earlier models read as Hopper
Models: DeepSeek-V4-Flash and MiniMax-M3 (B200/B300, 148 SMs); DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash (GB200/GB300, 152 SMs). Alibaba, Zhipu, Tencent, Xiaomi and StepFun models read as Hopper.
Method and evidence, as reported:
- PyTorch fills starting weights with the Philox generator via cuRAND, split across SMs × 2,048 threads. The same seed lays values out differently on 132-, 148- and 152-SM chips.
- Weights that training never touched (for example, unused-token rows) keep their starting values. Trained weights keep a weak correlation (r = 0.066, 57σ for V4-Flash's layers).
- Searched ~6 trillion seed × layout combinations, then tested the winning combination on held-out weights. Expected best chance match ≈8σ; reported matches 49–1,385σ, weakest single test 20σ.
- Hardware reproduction: seed 953049583 on a B300 reproduced 445 V4-Flash rows bit for bit (461/469 rows r>0.99); an H100 reproduced V3 but not V4-Flash; a rented GB200 reproduced all 13 preserved Kimi K3 rows (seed 20256586). Changing the seed by one broke every match.
- Agreement with all 24 publicly disclosed hardware cases. The author says these are not independent tests: 11 of them are step-zero checkpoints.
- Alternative ruled out for V4-Flash only: an AMD MI300X limited to 148/152 units reproduces the layout. A numerical-difference test pointed to NVIDIA in 363 of 377 values and to AMD in none.
Status unconfirmed: no independent public replication found as of Nov. 13, ~00:05 CT. Code and preserved rows are said to be public (repo not retrieved). Circumstantial support exists for DeepSeek (C-0056), Moonshot (C-0057) and Ant (GB300 benchmark logs, C-0069).
Sources
- S-0045 · Unemployment Labs (report) · 2026-11-12 · link (full)
- S-0044 · X (post + thread) · 2026-11-12 · link (full)
Related
- F-0042 Will DePue posts thread and report claiming Blackwell fingerprints in Chinese open-weight models
- F-0043 The DePue report states its own limits: initialization is not training, and location, ownership and legality are not established
- F-0044 The report lists no signature for DeepSeek-V4.1-Flash, in its main "Limits" section and appendix; the thread does not mention it
- C-0056 Unnamed senior administration official: DeepSeek’s upcoming model was trained on Blackwell, likely in Inner Mongolia
- C-0057 Kratsios: Moonshot "acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models"
Referenced by (15)
- claim C-0043 DePue: "conclusive forensic evidence" the top 4 Chinese labs built their newest models on Blackwell
- claim C-0045 DePue: the result was independently verified by senior researchers at OpenAI and Anthropic
- claim C-0055 Nvidia (Dec. 2025): "We haven’t seen any substantiation" of DeepSeek using smuggled Blackwells
- claim C-0056 Unnamed senior administration official: DeepSeek’s upcoming model was trained on Blackwell, likely in Inner Mongolia
- claim C-0057 Kratsios: Moonshot "acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models"
- claim C-0069 DePue report: a DeepSeek engineer wrote "we don’t have SM100 GPUs" (June 2025); Ant Group’s public repo has GB300 benchmark logs (Mar. 31, 2026)
- theory T-0001 Chinese intelligence operation (the published mainstream account)
- theory T-0003 Commercial chip-broker, not espionage
- theory T-0004 China-hawk faction and selective leaking
- theory T-0005 Anthropic-linked network (seed article’s theory)
- theory T-0006 Pro-Nvidia influence operation (counter-campaign)
- theory T-0007 US pretext for technology containment (PRC government position)
- theory T-0008 Governance and personal conduct (why he actually left)
- theory T-0009 Washington power struggle over Huang’s access to Trump
- theory T-0010 Nvidia’s China channel: diversion as business model, espionage as distraction