88Z AI, 20 November — Security firm Unemployment Labs introduces "Checkpoint Forensics," revealing that hardware-level differences in GPU architectures leave deterministic fingerprints in model weights. Variations in PRNG, CUDA kernel partitioning, and floating-point rounding between NVIDIA Hopper and Blackwell chips allow researchers to identify training infrastructure. Analysis indicates major Chinese open-source models shifted from Hopper to embargoed Blackwell hardware after April 2026.
At a glance
- GPU architecture differences leave deterministic hardware fingerprints in weight initialization
- Major Chinese open-source models shifted to Blackwell hardware after April 2026
- Public weight files may inadvertently reveal evidence of using embargoed hardware

Technical Principle: Hardware Fingerprints in Weights
Previously, speculation about the GPUs used by large model labs relied on supply chain rumors or insider leaks. A new report by security research firm Unemployment Labs (Will DePue’s team) introduces "Checkpoint Forensics." The core principle is that different GPU generations, such as NVIDIA Hopper (H100/H800) and the newer Blackwell (B200/GB200), have subtle but deterministic hardware-level differences in their underlying pseudo-random number generators (PRNG), CUDA kernel block partitioning, and floating-point rounding orders. These differences are embedded during weight random initialization at the start of training and subsequent tensor parallel sharding. Even after pre-training on trillions of tokens, residual initialization fingerprints in high-frequency spectra and specific tensor dimensions remain in the final open-source weight matrices like genetic markers.
Empirical Analysis: The Shift from Hopper to Blackwell
The research team relied solely on numerical reverse-engineering of publicly downloaded open-source model weights, validating findings by replaying computations on physical Blackwell and Hopper clusters. The results showed that mainstream Chinese open-source models released before April 2026 matched the NVIDIA Hopper architecture precisely. In contrast, six latest flagship open-source models released after April (including those from DeepSeek, Moonshot AI’s Kimi, MiniMax, and Ant Group) displayed numerical fingerprints consistent with NVIDIA Blackwell hardware, which is subject to strict export controls. This suggests that publishing safetensors weight files effectively provides evidence of the actual hardware configuration used in training, as every floating-point number bears the architectural signature.
88Z Perspective
This study highlights an overlooked security dimension in the open-source ecosystem: model weight transparency exposes not just algorithmic logic but also the physical attributes of infrastructure. For developers, this means "open sourcing" may inadvertently leak compliance-sensitive information, particularly regarding the use of export-controlled hardware. In the future, model publishers may need to consider noise injection or quantization before releasing weights to obscure hardware fingerprints, though this could impact model accuracy. For regulators and competitors, this offers a technical method to verify compute sources without infiltration. Practitioners should recognize that data privacy and hardware compliance face new scrutiny risks during model distribution.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- @mylifccSource published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.

