Weight study finds Blackwell fingerprints in six Chinese models
November 13 reporting added combined GPU estimates, core-count mappings, and the Kratsios intelligence claim; later coverage adds no new fingerprint findings.
A weight analysis links six Chinese models to Nvidia Blackwell chips but does not prove an export-control violation.
A forensic reading of published weights finds six models from DeepSeek, Moonshot AI, MiniMax and Ant Group match Nvidia Blackwell initialization patterns, while earlier models from those labs match Hopper. Former OpenAI researcher Will DePue maps DeepSeek-V4-Flash and MiniMax-M3 to B200 or B300 chips and four others to GB200 or GB300 systems. Hacker News estimates each run used 500 to 3,000 Blackwell GPUs for two to three months; The Decoder estimates five runs used about 12,000 combined. The authors say the fingerprint does not locate the chips or prove an export violation. TechCrunch says Blackwell exports to China are banned and that Michael Kratsios earlier cited US intelligence that Moonshot used banned chips.
- Evidence for the six models ranges from 49 to 1,385 sigma.
- The method matched known hardware for 24 control models.
- DeepSeek-V3's 2,000–2,500 chip estimate matches a disclosed 2,048.
- Two matches were confirmed on physical Blackwell chips.
- The Decoder puts Kimi K3 at about 4,500 Blackwell GPUs over 70 days.
- Examined Alibaba, Zhipu, Tencent, Xiaomi and StepFun models still read as Hopper.
What each outlet adds
beyond the shared facts- Hacker NewsSigma range, per-run GPU estimate, and two physical-chip confirmations.
- The Decoder148- and 152-core mappings and a 12,000-GPU combined estimate.
- TechCrunchPlaces DePue at Unemployment Labs and cites Kratsios on Moonshot.
Coverage
oldest first · one row per article · UTC- · 1d ago
Hacker News · securityFirst report
Six Chinese AI models match NVIDIA Blackwell GPU fingerprints
The analysis estimates each run needed roughly 500 to 3,000 Blackwell GPUs for two to three months.
- · 21h ago
The Decoder
Researcher says six Chinese models show Nvidia Blackwell fingerprints
Will DePue says unused weights preserve Hopper or Blackwell core-count patterns and published verification code.
- · 18h ago
TechCrunch · AI
Research claims banned Nvidia chips trained China’s leading open models
A former OpenAI researcher says Kimi K3 and DeepSeek V4 weights show Blackwell fingerprints; Jensen Huang left Nvidia this week.
- · 55m ago
Interconnects
Author says Huang exit endangers US open-model support
The essay argues Nvidia open-model funding will likely dry up as the company prioritizes stabilizing deliveries for 24 months.