> ## Content Index
> Fetch the complete content index at: https://claudefox.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Researcher says released weights show six Chinese models were started on Nvidia Blackwell GPUs
- URL: https://claudefox.com/checkpoint-forensics-chinese-models-blackwell-fingerprint/
- Published: 2026-11-13T10:08:35.000Z
- Updated: 2026-11-20T02:53:52.000Z
- Description: Will DePue's checkpoint-forensics report says the starting values left in published weights reveal which GPU created them, and that six models from DeepSeek, Moonshot, MiniMax and Ant Group match Nvidia Blackwell. How the method works, and what it cannot show.
- Author: hupicake
- Tags: Research

*Updated November 20, 2026: Expanded with how the method works, how to rerun the check yourself, and its stated limits.*

Will DePue of Unemployment Labs published a report on November 12 arguing that open model weights carry a fingerprint of the hardware that created them. He reads that fingerprint for dozens of public models. Six recent models from four Chinese labs, he says, were initialized on Nvidia Blackwell chips, which need a US export licence for China; the report notes that no such licences have been reported. US officials and press reports had already alleged Blackwell use by DeepSeek and Moonshot; what is new here is technical evidence anyone can check. The code to repeat the core check is public and runs without a GPU.

## What the report found

The six models are DeepSeek-V4-Flash and DeepSeek-V4-Pro, Moonshot's Kimi K3 and Kimi K3.1-Pro, MiniMax-M3 and Ant Group's Ling-3.1-flash, all released from April 2026 on. They are not all independent: K3.1-Pro starts from the same seed as K3 and may be a further-trained checkpoint of it. Earlier models from the same labs read as Nvidia's older Hopper generation: DeepSeek-V3, Kimi K2.5, MiniMax-M2.7 and Ling-3.0-flash. Kimi K2.7-Code sits in between: about 80% of its preserved rows carry the Hopper values inherited from K2.5 and about 20% match a 148-SM Blackwell layout, which the author reads as a merged checkpoint rather than mixed hardware, so it is not counted. The newest models he could read from Alibaba's Qwen, Zhipu's GLM, Tencent, Xiaomi and StepFun read as Hopper. Meituan's LongCat-2.0, which the company says was trained on ASIC superpods, showed no Nvidia fingerprint.

## How the fingerprint works

Before training, every weight needs a starting value. In PyTorch on Nvidia GPUs those values come from a deterministic random number generator, Philox, fed by a seed. The work is split across 2,048 threads per streaming multiprocessor (SM), and each thread writes to every Nth position, where N is the total thread count. The same seed therefore produces the same numbers on any chip, but in different places depending on how many SMs the chip has.

The SM counts separate the hardware families. Hopper H100, H800 and H200 have 132, so the method cannot tell them apart. Blackwell B200 and B300 have 148; Grace Blackwell GB200 and GB300 have 152\. Some weights, such as embedding rows for tokens never seen in training, keep their starting values almost exactly. The author first searched all 2^32 seeds against those rows under each candidate SM count to find each lab's seed and layout, then tested that choice on a separate set of weights that played no part in the search. The report says DeepSeek-V4-Flash and MiniMax-M3 match 148 SMs and the other four match 152\. Older chips cannot fake those layouts, since software can use fewer SMs than a chip has but not more, and the China-market Blackwell graphics cards produce neither total.

The report backs this up several ways. Its readings agree with all 24 models whose developers have named the hardware, such as Meta's Llama 3.1 on H100s and EleutherAI's Pythia on A100s, though the author notes that 11 are step-zero checkpoints and related models are not independent, so this is a consistency check rather than a measured accuracy rate. On physical chips, stock PyTorch with DeepSeek's seed reproduced all 57,344 preserved values of DeepSeek-V3 bit for bit on an H100, and 461 of 469 selected V4-Flash rows on a B300 (eight had been partly changed by training); neither chip reproduced the other model. A rented GB200 reproduced all 13 of Kimi K3's preserved rows, and changing the seed by one matched nothing. Even trained V4-Flash weights keep a faint trace under the Blackwell layout that disappears under Hopper. Statistically, the strongest evidence per model ranges from 49 sigma (V4-Pro, the weakest, where only part of the starting values survive) to 1,385 sigma (V4-Flash), against about 8 sigma expected from chance across roughly 6 trillion combinations.

## How to rerun the check

The repo macrogrants/checkpoint-forensics (MIT licence) ships the preserved rows, about 7 MB copied from Hugging Face and checked against pinned hashes, plus the recovered seeds (953049583 for DeepSeek, 20256586 for Kimi, 2760 plus the shard number for MiniMax, 3952 for Ling), so you do not need to repeat the brute-force search. replay.py rebuilds the rows for DeepSeek-V3, V4-Flash, Kimi K3, MiniMax-M3 and Ling-3.1-flash under every SM count from 16 to 256 and prints the one that matches; V4-Pro and K3.1-Pro are not in the replay set.

```
git clone https://github.com/macrogrants/checkpoint-forensics
cd checkpoint-forensics
pip install numpy scipy
python replay.py
```

A second script, gpu\_check.py, repeats the test with unmodified PyTorch on a GPU, so nothing in the repo generates the numbers. Rows only match on a chip with the right SM count. Setting REFETCH=1 downloads the rows again from Hugging Face.

## What it does not show

- Initialization, not training. The fingerprint identifies the chip that created the starting weights. A lab could create weights on one system and train elsewhere, so it supports, but does not prove, Blackwell training.
- Location and legality. The weights say nothing about where the chips were or who owned them. Blackwell systems can be rented outside China, and the report says it does not establish that export controls were broken.
- An AMD alternative. An AMD MI300X limited to 148 or 152 compute units can reproduce the same layouts. For DeepSeek-V4-Flash the author ruled this out numerically: of 377 values where the two vendors' generators differ, the weights match Nvidia in 363 and AMD in none. For the other five, the same test was not decisive and he only calls AMD very unlikely.
- Coverage. Closed models cannot be checked, and a blank result says nothing about hardware. DeepSeek-V4.1-Flash showed no signature after an exhaustive search; several Qwen3 models were initialized on the CPU; Baidu's ERNIE-5.5 uses PaddlePaddle, whose generator carries no chip layout.

The report also estimates the scale of the final training runs: about 12,000 Blackwell GPUs across five of them (roughly 30,000 H100 equivalents), with a range of 7,000 to 22,000\. Kimi K3 is the largest at about 4,500, then DeepSeek-V4-Pro at about 3,000; the others are near 1,500 each. These assume about 20% of peak GPU efficiency, estimated token counts and run lengths of 50 to 100 days, and that training stayed on the initializing hardware; halving run length doubles the count. The author says plainly that he is not confident in these numbers.

## Related on claudefox

- [llama-modchip runs the 284B DeepSeek-V4-Flash on a 64 GB PC with no GPU, using patched llama.cpp](https://claudefox.com/llama-modchip-deepseek-v4-flash-64gb-cpu/)
- [Edge0 open-sources a 145B model it says runs on a laptop; its own speed table uses RTX cards](https://claudefox.com/edge0-145b-open-model-local-hardware/)
- [rea-video: an open-source CLI that reverse-engineers a video shot by shot](https://claudefox.com/rea-video-reverse-engineer-video-shot-by-shot/)
- [A dragon that turns brush ink into watercolour](https://claudefox.com/a-dragon-that-turns-brush-ink-into-watercolour/)

## Our take

The strength of this work is that anyone can check the central claim in seconds, rather than taking a leak or an official statement on trust. The weak points are the ones the author flags himself: initialization is not training, the AMD test is only decisive for one model, and the GPU-count estimates are rough. For people who build with open weights, the practical point is wider than export policy: released checkpoints leak more about how they were made than most labs probably intend. Our guess, not the report's, is that some labs will now try to remove that signal.

Source: [Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs (Unemployment Labs)](https://unemploymentlabs.com/checkpoint-forensics?ref=claudefox.com), [macrogrants/checkpoint-forensics on GitHub](https://github.com/macrogrants/checkpoint-forensics?ref=claudefox.com)

This story first appeared in the [claudefox daily briefing for November 13, 2026](https://claudefox.com/daily-2026-11-13/).