---
type: analysis
slug: research-blackwell-fingerprint-claim-on-chinese-models-is-replayable-but-not-yet
title: "Blackwell fingerprint claim on Chinese models is replayable but not yet independently verified"
description: "Unemployment Labs' public code reproduces its own table for five checkpoints in seconds. No outside team has rerun the GPU tests, checked the AMD alternative, or examined the two \"Pro\" models, which the public repository does not include."
published_at: 2026-11-14T23:00:00.000Z
topics: ["trends","research"]
source_urls: ["https://unemploymentlabs.com/checkpoint-forensics","https://github.com/macrogrants/checkpoint-forensics","https://raw.githubusercontent.com/macrogrants/checkpoint-forensics/main/models.json","https://dev.to/frankchu/12000-banned-gpus-hid-in-chinas-open-weights-i-checked-1c65","https://the-decoder.com/former-openai-researcher-says-he-found-nvidia-blackwell-fingerprints-hidden-inside-six-chinese-ai-models/","https://techcrunch.com/2026/11/13/export-banned-nvidia-chips-are-behind-chinas-most-powerful-ai-models-uh-oh-jensen/"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
On November 12, 2026, Will DePue of the nonprofit Unemployment Labs published a checkpoint-forensics report. It argues that six open-weight Chinese models were initialized on export-controlled NVIDIA Blackwell GPUs: DeepSeek-V4-Flash and V4-Pro, Moonshot's Kimi K3 and K3.1-Pro, MiniMax-M3 and Ant Group's Ling-3.1-flash. The method relies on how PyTorch's CUDA random initialization works. Each value is assigned to a GPU thread in a way that depends on the chip's streaming-multiprocessor (SM) count. So embedding rows for tokens that training never touched still hold a layout that points to the chip family: 132 SMs for Hopper, 148 for B200/B300 and 152 for GB200/GB300.

This matters beyond export policy. If the fingerprint holds up, released weights can reveal training hardware, and labs can just as easily scrub it. The report says it does not establish smuggling or a legal violation. The question here is narrower: has anyone outside Unemployment Labs confirmed the technical result?

## What we checked

- **Can an independent team reproduce the Blackwell initialization fingerprints across the six models and validate the AMD alternative analysis?** partly confirmed. One outside developer reran the published replay.py on a CPU and got the author's table exactly. That replay covers only the five checkpoints in the repository, four of which the report reads as Blackwell, and it uses the author's chosen rows and seeds. We found no independent rerun on physical GPUs and no independent test of the AMD MI300X alternative. Two of the six models (DeepSeek-V4-Pro and Kimi K3.1-Pro) cannot be checked from the public artifacts. [4][2][3]
- **Do the public artifacts cover all six named models and the 24-model validation set?** refuted. models.json has five entries: DeepSeek-V3-Base (the Hopper baseline), DeepSeek-V4-Flash, Kimi-K3, MiniMax-M3 and Ling-3.1-flash-base. Neither V4-Pro, K3.1-Pro nor the 24 disclosed-hardware control models appear in it. [3][1]

## What we found

- The report's strongest single result is DeepSeek-V4-Flash. Unemployment Labs says unmodified PyTorch on a physical B300 regenerated 461 of 469 selected embedding rows. It also says its readings matched every case where a developer had disclosed its training hardware. Both are the author's own measurements. The list of 24 control cases appears in the report, but their preserved rows are not in the repository. [1]
- SM count alone does not rule out AMD. The report concedes that an MI300X can be partitioned to 152 or limited to 148 compute units. It separates the vendors with rounding differences, and that test is decisive only for V4-Flash. For the other five models it is the author's judgment, not a measurement. [1]
- The repository (MIT-licensed, a single commit when we opened it) does make the core claim checkable. replay.py regenerates preserved rows from pinned Hugging Face revisions with numpy and scipy. However, models.json lists only five checkpoints. DeepSeek-V4-Pro, Kimi K3.1-Pro and the 24 control models are absent, so the public tooling cannot test two of the six headline claims. [2][3]
- The only outside reproduction we found is a DEV Community post. Its author ran replay.py on a CPU and got the published table. That confirms the shipped code and data match what the report claims. It does not confirm the seed search, the GPU reruns or the AMD exclusion, because the replay starts from the author's recovered seeds and selected rows. [4]
- Independent coverage repeats the report's own caveats and adds no outside verification. The Decoder reports DePue's claim that unnamed senior researchers at OpenAI and Anthropic checked the method; none is named. TechCrunch and The Decoder report no response from DeepSeek, Moonshot AI, MiniMax, Ant Group or NVIDIA. The source that could overturn the main conclusion would be a lab statement or a failed independent GPU rerun. We searched for both and found neither. [5][6][1]

## What it means

For builders, the technique is credible and cheap to check. The geopolitical conclusion goes further than the independently checked evidence. What is verified so far is that the published code reproduces the published table for five checkpoints. The GPU reruns, the AMD exclusion, the two Pro models and the 24-case validation all rest on Unemployment Labs alone. Treat V4-Flash as the best-supported case and the other five as strong but single-source.

The practical lesson holds whichever reading proves right. Untouched embedding rows can leak the SM count, and likely the seed, of the machine that initialized a model. If you publish weights and consider your training hardware sensitive, re-randomize or prune unused vocabulary rows, or initialize on the CPU, before release. The report notes that DeepSeek-V4.1-Flash and several CPU-initialized models already show no fingerprint. If you audit third-party weights, the repository offers a ready starting point. Running gpu_check.py on your own B200/B300, GB200 or MI300X hardware would be a useful independent contribution.

## Competing readings

- **The fingerprints are real and the six models were initialized on NVIDIA Blackwell, as the report states; independent verification simply has not happened yet** (likely, about 70%). What would tell it apart: A B300 regenerated 461 of 469 V4-Flash rows, and the method matched all 24 disclosed cases [1]; Rounding differences favor NVIDIA over AMD for V4-Flash [1]; An outside CPU replay matched the published table [4].
- **V4-Flash is Blackwell, but some of the other five reflect AMD MI300X partitions or method error, which the evidence cannot yet exclude** (unlikely, about 25%). What would tell it apart: The MI300X can match the 148- and 152-unit layouts, and the AMD test is untested for five models [1]; Two of the six models are absent from the public manifest [3].
- **The method is flawed and the matches are artifacts** (very unlikely, about 5%). What would tell it apart: A seed+1 control and physical-GPU reruns would have to be wrong, yet the CPU replay is internally consistent; only the author has run the GPU checks [2][4].

## Base rate

Of 65 earlier research stories about companies, models, organisations or projects in the New Runtime archive over the last 12 months that are at least 3 months old, 47 (72%) were followed within 3 months by a release story on the same company or product line. In New Runtime's archive, 47 of 65 earlier research stories were followed within 3 months by a release story on the same entity. That tells us these labs will probably keep shipping. It says nothing about whether this forensic result will be independently verified.

## What would change this view

- An independent team runs gpu_check.py on its own B200/B300 or GB200 hardware and fails to reproduce the V4-Flash or Kimi-K3 matches (signal): The fingerprint would depend on something the author did, not on the hardware.
- An MI300X partitioned to 152 units reproduces the Kimi-K3 or Ling rows, including the rounding-level differences (signal): The AMD alternative would be live for those models.
- A lab publishes initialization logs or hardware disclosures that contradict the readings (model-listing signal): The method would conflict with ground truth, weakening the 24-case validation claim.

## Still open

- Will an independent team rerun gpu_check.py on its own Blackwell and Hopper hardware and get the same row matches?
- Can anyone run the vendor rounding test on an MI300X partitioned to 148 or 152 units for the five models where the AMD exclusion is untested?
- Will Unemployment Labs publish preserved rows and seeds for DeepSeek-V4-Pro, Kimi K3.1-Pro and the 24 control models?
- Will DeepSeek, Moonshot AI, MiniMax, Ant Group or NVIDIA respond on the record?

## Sources

1. [Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs (Unemployment Labs)](https://unemploymentlabs.com/checkpoint-forensics) (primary)
2. [macrogrants/checkpoint-forensics repository](https://github.com/macrogrants/checkpoint-forensics)
3. [checkpoint-forensics models.json](https://raw.githubusercontent.com/macrogrants/checkpoint-forensics/main/models.json)
4. [12,000 banned GPUs hid in China's open weights — I checked (DEV Community)](https://dev.to/frankchu/12000-banned-gpus-hid-in-chinas-open-weights-i-checked-1c65)
5. [The Decoder: Former OpenAI researcher says he found Nvidia Blackwell fingerprints](https://the-decoder.com/former-openai-researcher-says-he-found-nvidia-blackwell-fingerprints-hidden-inside-six-chinese-ai-models/)
6. [TechCrunch: Export-banned Nvidia chips are behind China's most powerful AI models](https://techcrunch.com/2026/11/13/export-banned-nvidia-chips-are-behind-chinas-most-powerful-ai-models-uh-oh-jensen/)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
