Agentic WeeklyGet The Weekly Verdict, freeSubscribe free

Tested Breaking

Report: DeepSeek, Kimi, MiniMax and Ling weights carry an export-controlled Nvidia Blackwell fingerprint

A Nov 12, 2026 report by Will DePue (Unemployment Labs) says the published weights of six models from DeepSeek, Moonshot AI, MiniMax and Ant Group were initialized on Nvidia Blackwell GPUs, which need a US license for China. It shows the chip class that made the starting weights, not where the chips were or that any law was broken.

By Agentic Weekly · Evidence: Tested · Editorial standards

Share on X
Facts listed
6
listed below, each from a source we read
We tested
Yes
we ran it; the notes are under \u201cWhat we have not verified\u201d
Gaps we flag
6
things we could not verify
Sources
4
2 primary, linked below

What happened

Facts the sources state, with the evidence status of each
#What the source saysStatus
01Reported. Will DePue's Nov 12 report says six models read as Blackwell-initialized: DeepSeek-V4-Flash and MiniMax-M3 (148 SMs, B200/B300) and DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash (152 SMs, GB200/GB300).Reported
02Reported. The same labs' earlier models read as Hopper (132 SMs): DeepSeek-V3, Kimi K2 and K2.5, MiniMax-M2.7 and Ling-3.0-flash. The report says Qwen, GLM, Tencent, Xiaomi and StepFun models it could read also show Hopper.Reported
03Reported. Method: PyTorch's GPU random-number fill lays out a seed's values by thread count (2,048 per SM), and embedding rows for tokens never seen in training keep those starting values. The report says it searched about 6 trillion seed and layout combinations.Reported
04Tested (CPU replay). We ran the author's public replay script: it rebuilt the untouched rows for five models and returned the claimed SM count for each, with seed+1 controls near zero. Rows we pulled from Hugging Face matched the repo's copies byte for byte.Reported
05Reported. The report estimates about 12,000 Blackwell GPUs across five final training runs (range 7,000 to 22,000) and says it is "not confident" in those figures; it revised them upward on Nov 12.Reported
06Not addressed. The report includes no response from DeepSeek, Moonshot AI, MiniMax, Ant Group, Nvidia or the Commerce Department, and we found none.Reported

reported by the sourcemeasured or run by usnot verified

Primary source: Unemployment Labs (Will DePue): Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs (Nov 12, 2026). On-the-record statements and letters are the source of record; press reports are named and attributed where we use them.

What we checked vs what is the author's claim

Tested here means one narrow thing: the public replay reproduces the author's numbers on rows that really are in the published Hugging Face weights. It does not mean we confirmed which hardware trained the models. The SM-to-chip mapping, the GPU runs and the hardware conclusions are the author's.

Each claim, its label and who says so (sources read Nov 13, 2026)
ClaimLabelWho says so
Untouched embedding rows in five models match one seed at 132, 148 or 152 SMs, and seed+1 matches nothingTested (CPU replay)Our run of macrogrants/checkpoint-forensics; also peterjliu/checkpoint-forensics-repro
148 SMs means B200/B300 and 152 means GB200/GB300ReportedReport; the third-party repro says it did not check this and that public spec pages disagree
A B300 reproduced DeepSeek-V4-Flash's rows bit for bit and a rented GB200 reproduced Kimi K3's; an H100 reproduced none of V4-Flash'sReportedReport and repo README
AMD MI300X in two-way partition mode also has 152 compute units; excluded for V4-Flash, called very unlikely for the restReportedReport
About 12,000 Blackwell GPUs and over half a billion dollars across the runsReported estimateReport, which says it is not confident in it
Senior OpenAI and Anthropic researchers independently verified the resultUnverifiedThe author's X thread; no reviewers are named

What this means for export controls (our reading)

If the fingerprint holds, it is the first public evidence drawn from the labs' own released files rather than leaks or official statements. US officials and press reports had already alleged Blackwell use by DeepSeek and Moonshot, according to the report, but without technical evidence anyone could re-run.

It does not settle the legal question. Weights can be created on rented chips outside China, and the report says itself that its findings do not establish where the chips were, how access was obtained, or whether export controls were violated. The story lands the day after Democrats opened a probe into Nvidia H200 export licenses, so expect it to be cited in that fight.

Why it matters if you build on these models

Nothing changes in the weights you download, and the report makes no claim about model quality or safety. The practical risk is policy: open-weight models that become the subject of export-control or procurement rules can be restricted for some customers, so teams in regulated industries may want to note which models in their stack are named.

Hard questions still unanswered

  • Where were the Blackwell systems physically located, and who owned or rented them?
  • Did training stay on the hardware that created the starting weights, or only initialization?
  • Will any of the four labs, or Nvidia, say what hardware these runs used?
  • Will someone with GB200 access re-run the Kimi K3 and DeepSeek-V4-Pro checks on stock PyTorch, as the repo's gpu_check.py allows?

In the source’s words

“The fingerprint identifies how the weights were created and supports, without proving, the inference that Blackwell also performed the training.”

“Notably, our findings provide technical evidence of access to advanced computing hardware, but do not establish where the chips were located, how access was obtained, or whether export controls were violated.”

“It does not show where the chips were, who owned them, or that the whole training run used them.”

What we have not verified

  • We did not run anything on a GPU; the B300, GB200 and H100 results are the author's.
  • Which chips have 148 and 152 SMs is taken from the report; we did not check it on hardware.
  • We did not check DeepSeek-V4-Pro or Kimi K3.1-Pro, the 24 known-hardware cases, or the AMD exclusion.
  • The GPU-count and cost estimates rest on the author's assumptions about efficiency, tokens and run length, which no lab has published.
  • The claim that OpenAI and Anthropic researchers verified the result comes from the author's X thread; no reviewers are named.
  • No lab, Nvidia or the Commerce Department has responded in what we read.

Tested: On Nov 13, 2026 we cloned github.com/macrogrants/checkpoint-forensics and ran its CPU replay (python replay.py, numpy 2.5.4, scipy 1.18.1, about 4 seconds). It returned the report's layouts: DeepSeek-V3-Base 132 SMs (8/8 rows), DeepSeek-V4-Flash 148 (461/469), Kimi-K3 152 (13/13), MiniMax-M3 148 (17/17), Ling-3.1-flash-base 152 (384/384, r 0.983); seed+1 controls stayed near 0. We then pulled 20 of those rows straight from the pinned Hugging Face revisions and they were byte-identical to the repo's copies. Not covered: no GPU run, DeepSeek-V4-Pro and Kimi K3.1-Pro (not in the replay), the 24 known-hardware cases, the AMD exclusion, which chip has which SM count, and the GPU-count and cost estimates.

Sources

Spotted an error? See our corrections policy or email the address on the contact page. Markdown version for agents: /news/chinese-ai-models-blackwell-gpu-fingerprint.md

Want these briefs in your Google Top Stories? Add Agentic Weekly as a preferred source on Google.

Newsletter

The Weekly Verdict. Free, every Monday.

A free email every Monday: the week's few big AI releases, then 5 short use cases for the new models, each with a prompt you can paste.

Free. Double opt-in. One click to unsubscribe.