To investigate hardware traces in public model weights, start with the forensic report and checking code released by Unemployment Labs on November 12, 2026. Before choosing between a numerical check on a laptop and a rerun on a physical GPU, establish whether the method applies to your target model.
In the report's verification section, Will DePue says unmodified PyTorch running on a B300 reproduced 445 complete rows of DeepSeek-V4-Flash weights bit for bit. That is a reproduction count, not something readers can automatically treat as the number of independent statistical observations.
Published weights are only one prerequisite. They must also retain a recognizable, hardware-dependent pattern from the initialization software. The report says several Qwen3 models were initialized on CPUs and therefore carry no fingerprint of the GPUs used for training. An absent match does not establish that a particular GPU family was never used.
Start with the numerical check before arranging a GPU rerun
The entry point is the author's code repository. Its replay.py reconstructs values from the seed and compares them with saved weight rows, without requiring a GPU. The author says it runs on a laptop in a few seconds; that is not our timing result. By contrast, gpu_check.py uses unmodified PyTorch to generate starting values on a physical GPU, testing whether that device reproduces the result. The two checks serve different purposes.
For a replication attempt, record the checkpoint version, selected rows, random seed, software versions and device configuration. Save the controls, not just the successful output. The author's checks include changing the seed and switching between Hopper and Blackwell chips. He reports that neither device reproduced the preserved rows associated with the other model.
The PyTorch reproducibility documentation does not promise identical results across releases, commits or platforms. A matching seed is therefore not a guarantee that every environment should produce the same output. Equally, a failed attempt is not enough to reject the report before checking whether the environments correspond.
An unusual match still needs an explanation
The report uses sigma to describe how far a match exceeds accidental resemblance, conditional on its statistical assumptions. Hardware identification is a separate part of the investigation. A specially configured AMD MI300X can reproduce the relevant layouts, according to the report. Readers need the additional tests for a particular model to assess alternatives; choosing the largest score does not do that work.
Weights can move to different equipment after initialization. Their numerical patterns do not disclose the equipment's location or how access was obtained. If your decision is simply whether to use a model, evaluate quality, latency and cost on your own tasks. If you need to verify its training provenance, ask for reproducible configurations and training records. Those are different decisions, and a single significance number cannot answer both.