Six Chinese AI models show NVIDIA Blackwell GPU fingerprints
A forensic report identifies NVIDIA Blackwell GPU fingerprints in six models from DeepSeek, Moonshot AI, MiniMax and Ant Group. The evidence comes from public model weights, not leaks or disclosures. It gives regulators and users a technical method to trace restricted hardware use in frontier AI development.
Unemployment Labs analyzed publicly released model weights and found six frontier models from four Chinese AI developers contain numerical fingerprints matching NVIDIA Blackwell GPUs. The models include DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash. The report says the fingerprints identify how initial weights were created and support the inference that Blackwell hardware was used. It does not prove the full training run occurred on those chips.
The fingerprint method recovers initial values from model checkpoints and searches for hardware and software configurations that reproduce them. DeepSeek-V4-Flash and MiniMax-M3 match B200 or B300 chips with 148 streaming multiprocessors. DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash match GB200 or GB300 chips with 152 streaming multiprocessors. The report says DeepSeek-V4-Flash and Kimi K3 were also checked on physical Blackwell chips. Specially configured AMD hardware is the only alternative found, excluded for DeepSeek-V4-Flash and considered very unlikely for the others.
The timing places Blackwell use in 2026. DeepSeek’s models read as Hopper through DeepSeek-V3 in December 2024 and as Blackwell from April 2026. Moonshot’s models read as Hopper through Kimi K2.5 in January 2026 and as Blackwell from July 2026. MiniMax shifted by June 2026. Ant Group shifted by October 2026. The Hopper readings do not identify a specific chip; H800s, H100 SXMs and H200s share the fingerprint. The report does not establish chip locations, procurement routes, or whether export controls were violated.
The finding matters because U.S. export controls target access to Blackwell-class accelerators, which are central to frontier AI training. Traceable models give enforcement and supply-chain audits a verifiable artifact independent of corporate disclosure. The report also complicates claims that Chinese labs rely only on domestic or older chips. It affects chipmakers, regulators, model users and investors by making hardware provenance a testable part of AI model evaluation.
This affects chipmakers, regulators and AI users by turning restricted GPU use into a checkable technical fact inside model weights. It gives enforcement and supply-chain audits evidence that does not depend on company disclosure.
Sources (1)
- [1]Original source(https://unemploymentlabs.com/checkpoint-forensics)