Will DePue of Unemployment Labs says six open-weight AI models from DeepSeek, Moonshot AI, MiniMax and Ant Group carry numerical fingerprints matching Nvidia’s Blackwell GPUs. His report, published November 12, 2026, bases that claim on the models’ publicly released weights, with additional checks on physical Blackwell hardware for two models.
The distinction that matters is initialization versus training. DePue’s method attempts to identify the hardware configuration that generated a model’s starting weights. The report argues that this supports an inference about training hardware, but it cannot prove where the chips were located, how access was obtained, whether all training ran on Blackwell, or whether any law was broken.
That narrower claim is still worth examining. Public model weights could offer an inspectable source of evidence about advanced computing access, separate from company disclosures, intelligence claims or supply-chain reporting. The useful question is how much of that history the numbers can actually reveal.
How Model Weights Can Retain a Hardware Fingerprint
An AI model’s weights are numerical values that training repeatedly adjusts. Before that process begins, many weights receive pseudorandom starting values generated from a seed.
A seed makes the process reproducible under matching conditions. It is not enough, however, to know only the seed. According to DePue’s explanation, the way software distributes initialization work across a GPU also affects which generated numbers land in which positions.
His method relies on three connected observations:
- Initialization is structured. PyTorch distributes the work across many parallel threads, each generating and placing part of the numerical sequence.
- Hardware layout affects that distribution. The number of streaming multiprocessors, or SMs, influences the resulting arrangement.
- Training can leave traces behind. Some weights remain unchanged, while others retain a weaker statistical relationship with their starting values.
DePue says those remnants let him work backward: choose a candidate seed and hardware layout, generate the corresponding starting values, and compare them with a released checkpoint.
The Unemployment Labs report distinguishes relevant Hopper configurations with 132 SMs from Blackwell configurations with 148 or 152 SMs. These are streaming-multiprocessor counts, not counts of individual arithmetic cores.
The proposed fingerprint is therefore an interaction between the initialization software, seed and hardware configuration. It is not a serial number embedded by Nvidia, and it is not an intentional disclosure by the model developer.
That distinction helps explain both the method’s appeal and its limits. If the same configuration reliably recreates surviving values, researchers have something more concrete than an architectural resemblance. But identifying a numerical layout still requires checking whether another software or hardware setup could produce it.
The Six Models Named in the Report
Unemployment Labs identifies the following six releases as carrying Blackwell-matching initialization patterns. The release months and chip-family matches below come from the report’s comparison table.
| Developer | Model | Released | Reported initialization match |
|---|---|---|---|
| DeepSeek | DeepSeek-V4-Flash | April 2026 | B200/B300, 148 SMs |
| DeepSeek | DeepSeek-V4-Pro | April 2026 | GB200/GB300, 152 SMs |
| Moonshot AI | Kimi K3 | July 2026 | GB200/GB300, 152 SMs |
| Moonshot AI | Kimi K3.1-Pro | November 2026 | GB200/GB300, 152 SMs |
| MiniMax | MiniMax-M3 | June 2026 | B200/B300, 148 SMs |
| Ant Group | Ling-3.1-flash | October 2026 | GB200/GB300, 152 SMs |
The paired names matter. A B200/B300 match does not distinguish those two products, and a GB200/GB300 match does not distinguish between those configurations. The report’s inference is more specific than “some Nvidia GPU,” but less specific than identifying an exact purchased system.
DePue also reports a generational pattern within each developer’s releases:
- DeepSeek: DeepSeek-V3 matches Hopper; the V4 models match Blackwell.
- Moonshot AI: Kimi K2 and K2.5 match Hopper; K3 and K3.1-Pro match Blackwell.
- MiniMax: M2.7 matches Hopper; M3 matches Blackwell.
- Ant Group: Ling-3.0-flash matches Hopper; Ling-3.1-flash matches Blackwell.
The report cautions that its Hopper fingerprint cannot distinguish an H800 from an H100 SXM or H200. Where its table names H800 hardware for earlier models, those identifications come from developer disclosures, not from the fingerprint alone.
Six Releases Do Not Necessarily Mean Six Training Runs
Unemployment Labs says Kimi K3.1-Pro starts from the same seed as K3 and may be a K3 checkpoint with additional post-training.
That is a consequential detail. A later release can inherit an initialization fingerprint through its parent checkpoint. It need not represent a new model initialized from scratch on another cluster.
I would count these as six reported model matches, not six demonstrated independent Blackwell pretraining runs. The shared seed supports a possible relationship between the two Kimi releases, while the report itself leaves that relationship qualified.
The same caution applies to timing. Release dates show when checkpoints became public; they do not establish the date on which a developer first obtained access to a particular GPU.
What the Physical Blackwell Checks Show
DePue says he went beyond searching candidate numerical layouts for two models. In his public explanation of the work, he describes running unmodified PyTorch on physical Blackwell hardware using recovered seeds.
His reported results differ in an important way:
| Model | Physical hardware | Reported result |
|---|---|---|
| DeepSeek-V4-Flash | Nvidia B300 | Untouched starting weights reproduced bit for bit |
| Kimi K3 | Rented Nvidia GB200 | Correlation of 0.999995 with recovered starting-weight traces |
For DeepSeek-V4-Flash, DePue also says an H100 reproduced none of the relevant untouched starting weights, and changing the recovered seed by one broke the match.
Those controls make the claim more specific. A successful candidate configuration, an unsuccessful alternative GPU and a failed nearby seed provide different checks on the proposed explanation.
The two successful results should not be flattened into the same description, though. Bit-for-bit agreement is exact agreement on the values being compared. A correlation of 0.999995 describes an extremely close statistical relationship, not identity across every value.
Neither result means that DePue reproduced an entire trained model. The claimed reproduction concerns initialization traces recovered from released weights.
My view is that these physical checks are the most useful part of the evidence to prioritize for replication. They connect the search result to real hardware and a named software framework, giving another researcher a more concrete target than a broad statement that a model “looks Blackwell-trained.”
How Strong Is the Statistical Case?
Unemployment Labs says it searched roughly 6 trillion seed-and-layout combinations. A search that large needs to account for accidental matches: trying enough candidates can make unrelated data appear meaningful.
According to the report’s calculations, the strongest chance match across that search would be expected to reach about 8 sigma. The strongest evidence for each of the six models ranges from 49 to 1,385 sigma, while even the weakest individual test it reports reaches 20 sigma.
These are DePue’s statistical results. They should not be translated into a percentage probability that a company used Blackwell, or treated as a probability that an export-control violation occurred.
The distinction is between two questions:
- Could unrelated numerical values produce a match this strong under the report’s search assumptions?
- Could another real initialization setup produce the same pattern?
A large statistical separation can address the first question without settling every possible explanation in the second.
Unemployment Labs says its method also agrees with developer-disclosed hardware for 24 models with known answers. That is a useful cross-check because it tests the method against information outside the six disputed cases.
DePue further says senior researchers at OpenAI and Anthropic independently checked his findings, a claim also reported by The Decoder. The cited material does not provide a separately documented replication of the full forensic analysis, so those checks remain described through his account.
The AMD Alternative Shows Why Configuration Matters
Frequently Asked Questions
3 questions
1What Is a Weight-Initialization Fingerprint?
It is a numerical pattern left by the process that generated a model’s starting weights. DePue says some of that pattern survives training and can be matched to candidate seeds and GPU configurations.
