Skip to content
Skip to the story
Dark watercolour semiconductor plate backdrop for the checkpoint forensics report

A folkfox special report.

Checkpoint forensics: the six hidden Blackwell clues inside open weights

Summarise with
Set as preferred source on Google
17 min read

A PyTorch CUDA quirk turns GPU randomness into hardware trails. On 12 November 2026, Will DePue published checkpoint forensics matching six Chinese open models to NVIDIA Blackwell grids.

By Katie Delaney. Research checked 13 November 2026. The illustrations are conceptual.

folkfox vixen inspecting a dark silicon wafer with a brass loupe by lantern light for checkpoint forensics

How checkpoint forensics reads hidden chip layouts

Neural weights hold a faint hardware scent long after pretraining ends. On 12 November 2026, Will DePue published checkpoint forensics findings at Unemployment Labs showing open weights retain GPU thread signatures.

Checkpoint forensics exploits routine software determinism rather than hidden watermarks. Before step one, PyTorch lays random seeds across GPU threads by processor count, shifting initial weights whenever one streaming multiprocessor changes.

X Thread: How Will DePue framed checkpoint forensics

Will DePue@willdepue
Summary of X post

Matches six Chinese models across six trillion trials to 148-SM and 152-SM Blackwell grids, reproducing DeepSeek V4 Flash untouched weights bit for bit on a B300 at seed 953049583.

12 Nov 2026, 21:29 UTC · 14,253 likes · 1,949 repostsView post on X
Will DePue@willdepue
Excerpt

The random numbers a model starts with aren’t truly random. They come from a formula that starts from one starting number called the seed, and using the same seed gives the same numbers every time. [...]

[...] Hopper has 132, Blackwell has 148 and Grace Blackwell has 152. So the same seed lays the numbers out differently on each chip. [...]

12 Nov 2026, 21:29 UTC · 402 likes · 13 repostsView post on X

How checkpoint forensics tracks CUDA Philox grids

Calling torch.manual_seed on NVIDIA GPUs runs PyTorch ATen CUDA kernels using cuRAND Philox4x32-10 and Box–Muller sampling.

Because PyTorch assigns 256-thread blocks up to 2,048 threads per SM, documented in reproducibility notes and CUDA RNG issues, thread stride separates Hopper (132 SMs), NVIDIA B200/B300 (148 SMs) and Grace Blackwell (152 SMs).

folkfox technical diagram

How PyTorch CUDA thread count changes the weight grid

torch.manual_seed(s) -> Philox4x32-10 + curand_normal4 (256 threads/block)NVIDIA Hopper (H100 / H800 / H200)132 SMs x 2,048 threads/SM (8 blocks x 256)Stride: 270,336NVIDIA Blackwell (B200 / B300)148 SMs x 2,048 threads/SM (+32,768 vs Hopper)Stride: 303,104Grace Blackwell (GB200 / GB300)152 SMs x 2,048 threads/SM (+8,192 vs B200)Stride: 311,296
How PyTorch CUDA thread count changes the weight grid
TierSMsThreadsModels
NVIDIA Hopper (H100 / H800 / H200)132 SMs270,336 threadsDeepSeek-V3, MiniMax-M2.7, Ling-3.0-flash
NVIDIA Blackwell (B200 / B300)148 SMs303,104 threadsDeepSeek V4 Flash, MiniMax M3
Grace Blackwell (GB200 / GB300)152 SMs311,296 threadsV4-Pro, Kimi K3, K3.1-Pro, Ling-3.1-flash
Identical seeds populate shifted tensor indices across Hopper, B200/B300 and GB200/GB300 thread grids.

Source: PyTorch DistributionTemplates.h, NVIDIA cuRAND docs and Unemployment Labs, 13 Nov 2026.

Why checkpoint forensics finds cold trails

Training never scrubs every row. Cold MoE experts stay frozen and weight decay keeps residual signal, extending DeepSeek’s runtime trails.

In control runs, DePue painted a smiley face and portrait inside starting weights; both survived pretraining, showing how checkpoint forensics detects layout geometry.

Matt Thompson@fuzziphy
Verbatim

@willdepue @aidenybai GPUs use a Philox PRNG.. so the layout of SM thread/block/warp leaves a geometric signature matching the kernel layout?

12 Nov 2026, 23:29 UTCView post on X
Pedro G@pggalaviz
Verbatim

@willdepue A decade of everyone cursing that you can't get bit-for-bit reproducible results across GPUs, and that exact quirk ends up being the evidence. Nondeterminism finally found a job.

12 Nov 2026, 23:24 UTC · 4 likesView post on X
folkfox vixen consulting twin ledger dials under cold moonlight representing GPU cluster statistics

Six models matched from DeepSeek V4 Flash to MiniMax M3

Scanning six trillion trials via checkpoint forensics splits older Hopper stacks from 2026 Blackwell arrivals. DeepSeek V4 Flash and MiniMax M3 match 148-SM layouts shared by NVIDIA B200 and B300 silicon.

Four flagship models match 152-SM Grace Blackwell GB200/GB300 grids: DeepSeek-V4-Pro, Moonshot AI’s Kimi K3 and Kimi K3.1-Pro, and Ant Group’s Ling-3.1-flash. Earlier Kimi K2.7-Code blends 132 and 148 SMs (80/20).

DeepSeek V4 Flash and Kimi K3 matches via checkpoint forensics

On DeepSeek V4 Flash (284B total, 13B active, April 2026), checkpoint forensics scores 1,385 sigma: 445 of 469 zero-update rows match B300 identically, 461 top r = 0.99, and 745,472 trained weights retain r = 0.066 (57 sigma).

On rented GB200 silicon, seed 20256586 matched all 13 untouched Kimi K3 rows at r = 0.999995. H100 matched zero DeepSeek V4 Flash rows, while 24 controls matched stated GPUs.

folkfox chart

Checkpoint forensics Z-scores vs the 8-sigma random search ceiling

Checkpoint forensics Z-scores vs the 8-sigma random search ceiling8 sigma chance ceiling (6 trillion trials)DeepSeek-V4-Flash1,385 sigma · 148 SMMiniMax-M3323 sigma · 148 SMKimi K3305 sigma · 152 SMKimi K3.1-Pro305 sigma · 152 SMKimi K2.7-Code246 sigma · 80/20 blendLing-3.1-flash244 sigma · 152 SMDeepSeek-V4-Pro49 sigma · 152 SM
Checkpoint forensics Z-scores vs the 8-sigma random search ceiling
ModelBest Z-scoreMatched grid
DeepSeek-V4-Flash1385 sigma148 SM (B200/B300)
MiniMax-M3323 sigma148 SM (B200/B300)
Kimi K3305 sigma152 SM (GB200/GB300)
Kimi K3.1-Pro305 sigma152 SM (GB200/GB300)
Kimi K2.7-Code246 sigma80/20 132+148 blend
Ling-3.1-flash244 sigma152 SM (GB200/GB300)
DeepSeek-V4-Pro49 sigma152 SM (GB200/GB300)
Across six trillion trials, random noise peaks near 8 sigma; the six flagged Blackwell releases score between 49 and 1,385 sigma.

Source: Unemployment Labs Checkpoint Forensics (12 November 2026).

MiniMax M3, Ant Group logs and empty controls

Public papers blurred MiniMax and Ant Group's chip trails. While MiniMax M3 shipped on 1–2 June 2026 (428B / 23B active), MiniMax’s June MSA paper tested H800 speeds on a separate 109B model while M3 shifted to 148-SM Blackwell (323 sigma).

Ant’s InclusionAI shifted from July’s 132-SM Ling-3.0-flash to Ling-3.1-flash (560B / 25B active, 244 sigma at 152 SMs). Ant’s cuLA GB300 log carries a 31 March 2026 timestamp.

X Thread: Model spikes and MiniMax/Ant shift

Will DePue@willdepue
Excerpt

[...] This is that search on DeepSeek-V4-Flash. Every processor count gives noise until 148 (Blackwell), where the weights match.

12 Nov 2026, 21:29 UTC · 363 likes · 15 repostsView post on X
Will DePue@willdepue
Verbatim

Critically, MiniMax and Ant have not been publicly tied to Blackwell before. MiniMax’s models match Hopper through M2.7 in April and Blackwell from M3 in June. Ant’s own GitHub has benchmark logs from a GB300 dated March 31.

12 Nov 2026, 21:29 UTC · 303 likes · 12 repostsView post on X

Empty controls prove equally instructive. Checkpoint forensics finds no CUDA Philox trace in DeepSeek-V4.1-Flash, ERNIE-5.5 (PaddlePaddle seed 2760), Qwen3 (CPU init) or LongCat-2.0, so viral X claims treating V4.1-Flash as proven Ascend training exceed this negative test.

2026 Chinese open-weight models tested by checkpoint forensics vs public technical records
Model & labRelease windowMatched grid & sigmaCorroboration & caveat
DeepSeek V4 Flash (DeepSeek)Apr 2026 (284B / 13B)148 SM (B200/B300), 1,385 sigma445/469 exact rows on B300; 363/377 NVIDIA vs 0 AMD
DeepSeek-V4-Pro (DeepSeek)Apr 2026152 SM (GB200/GB300), 49 sigmaWeakest flagship score; AMD edge case unresolved
Kimi K3 & K3.1-Pro (Moonshot)Jul & Nov 2026152 SM (GB200/GB300), 305 sigma13/13 Kimi K3 rows on GB200 at r = 0.999995
Kimi K2.7-Code (Moonshot)Jun 202680/20 Hopper+148 SM, 246 sigmaBlended Hopper-to-Blackwell transition release
MiniMax M3 (MiniMax)Jun 2026 (428B / 23B)148 SM (B200/B300), 323 sigmaM2.7 matched 132 SM; MSA H800 benchmark used 109B model
Ling-3.1-flash (Ant Group)Sep–Oct 2026 (560B / 25B)152 SM (GB200/GB300), 244 sigmaLing-3.0-flash matched 132 SM; cuLA log dated 31 Mar
DeepSeek-V4.1-Flash & Qwen3Control suiteEmpty GPU fingerprintNon-CUDA or CPU init proves neither NVIDIA nor Ascend
folkfox vixen examining an illuminated analytical scroll surrounded by glass calibration instruments

Kernels, scale and five critical gaps

Within hours, researchers debated on X where checkpoint forensics stops and speculation begins. Some pointed directly to GitHub kernels; statisticians stressed five inferential gaps between seed maths and customs proof.

SemiAnalysis founder Dylan Patel noted Moonshot and DeepSeek already shipped Blackwell kernels, echoed by @ariaurelium. DePue replied that public customer kernels differ from training models on restricted silicon.

X Exchange: Open-source kernels vs training clusters

aria@ariaurelium
Verbatim

this seems like a lot of effort to forensically prove something that nobody is concealing

you can just look on github and see that DeepSeek regularly publishes Blackwell training/inference kernels

12 Nov 2026, 21:54 UTC · 88 likes · 3 repostsView post on X
Will DePue@willdepue
Summary of X post

Separates releasing public Blackwell inference kernels from financing billion-dollar training runs on restricted silicon.

12 Nov 2026, 21:56 UTC · 47 likesView post on X

Five methodological limits in checkpoint forensics

Jonathan Sandhu praised the physical B300 and GB200 reproductions while listing five unresolved gaps:

  • Init hardware split from full pretraining cluster
  • Partitioned AMD MI300X ambiguity outside DeepSeek V4 Flash
  • Zero 152-SM separation across GB200 and GB300
  • Modelled 2k–3k chip estimates assuming 35% utilisation
  • Private lab checks lacking public scripts

Sandhu sought blind cross-chip and partitioned AMD tests with open code. Analyst @teortaxesTex noted 2,000–3,000 Blackwells match sub-10MW clusters, fitting the MoE scales below.

Jonathan Sandhu@SandhuJonathan
Summary of X post

Separates sound Philox engineering from procurement claims: init leaves location unproven, MI300X/GB300 overlap remains open, and $500M+ assumes 35% utilisation.

12 Nov 2026, 23:42 UTCView post on X
Teortaxes@teortaxesTex
Summary of X post

Notes 2,000–3,000 Blackwells imply sub-10MW clusters, while agreeing checkpoint forensics confirms active Blackwell stacks.

12 Nov 2026, 22:09 UTC · 92 likes · 7 repostsView post on X

folkfox chart

Model scale across the 2026 Hopper-to-Blackwell shift (billions)

Model scale across the 2026 Hopper-to-Blackwell shift (billions)Blackwell (148/152 SM)Hopper (132 SM)Active params (B)Ling-3.1-flash152 SM · GB200/GB300560B / 25B activeMiniMax M3148 SM · B200/B300428B / 23B activeDeepSeek V4 Flash148 SM · B200/B300284B / 13B activeMiniMax-M2.7132 SM · Hopper229B / 10B activeLing-3.0-flash132 SM · Hopper124B / 5.1B active
Model scale across the 2026 Hopper-to-Blackwell shift (billions)
ModelTotal parametersActive per tokenMatched layout
Ling-3.1-flash560B25B152 SM · GB200/GB300
MiniMax M3428B23B148 SM · B200/B300
DeepSeek V4 Flash284B13B148 SM · B200/B300
MiniMax-M2.7229B10B132 SM · Hopper
Ling-3.0-flash124B5.1B132 SM · Hopper
Verified Mixture-of-Experts footprints explain why 13B–25B active models fit sub-10MW 2k–3k chip runs while total capacity climbs sharply.

Source: Hugging Face official model cards and Unemployment Labs SM grids.

Neither DeepSeek, Moonshot AI, MiniMax, Ant Group, NVIDIA, OpenAI nor Anthropic posted public replies on X by midnight UTC, and independent checkpoint forensics replications remain pending. Our NVIDIA enterprise infrastructure checks explain why open code matters alongside weights.

Forensic evidence confidence vs unresolved inferential gaps across flagged checkpoints
DimensionObserved evidenceUnresolved gapConfidence level
Initialisation match445/469 V4-Flash rows exact on B300; 13/13 Kimi K3 on GB200Public repo pending; scripts currently by requestVery high (V4-Flash & Kimi K3)
NVIDIA vs AMD MI300X363/377 NVIDIA cuRAND vs 0 AMD matches on V4-FlashValue-level AMD test unconfirmed for all 5 peersHigh (V4-Flash); moderate elsewhere
Full training runResidual correlation r = 0.066 (57 sigma) in trained layersCross-chip transfer after init unruled out empiricallyStrong inference, not proof
Cluster count & spendModelled ~2k–3k Blackwells & $500M+ across labsAssumes 35% MFU utilisation and training cadenceModelled estimate only
Physical locationSix models carry 148/152-SM Blackwell signaturesWeights cannot tell smuggled chips from foreign cloudsZero direct visibility
folkfox vixen standing beside an amber harbour lantern watching fog roll past regulatory charts

NVIDIA B200, offshore clouds and the Chip Security Act

Export rules explain why checkpoint forensics triggered immediate Washington concern. Four years of US controls restricted mainland chip sales while offshore cloud rentals and smuggling opened a regulatory thicket.

Under updated US Bureau of Industry and Security (BIS) rules, top GPUs stay under tight export-licence limits. A narrow January 2026 exemption allowed capped H200 sales below 21,000 TPP and 6,500 GB/s.

Where NVIDIA B200 and GB300 sit in export rules

NVIDIA B200, B300, GB200 and GB300 exceed those thresholds and remain banned for mainland sale. Yet Carnegie Endowment research (May 2026) tracks Chinese labs renting legal NVIDIA B200 clouds in Malaysia, Thailand and Japan.

Checkpoint forensics still leaves geography open: May 2026 BIS guidance exempted pre-existing overseas centres, while DOJ smuggling cases totalled $160M, $300M and $2.5B.

Will DePue@willdepue
Verbatim

Of course, this tells us nothing about where these chips physically are. They could have been smuggled into China or the labs are using them abroad through gaps that Congress has not closed. In either case, it’s clear that the export bans are deeply failing.

12 Nov 2026, 21:29 UTC · 392 likes · 18 repostsView post on X
Dean W. Ball@deanwball
Verbatim

When I first saw this I thought, “wow, this is dna evidence,” and then it quickly dawned me that this is in fact orders of magnitude stronger than literal dna evidence submitted in court.

12 Nov 2026, 22:09 UTC · 198 likes · 11 repostsView post on X
Patrick McKenzie@patio11
Verbatim

One of the more interesting bits of technical independent journalism (with shades of security research) I’ve ever read, and it has obvious policy implications if accurate.

12 Nov 2026, 22:31 UTC · 202 likes · 13 repostsView post on X
roon@tszzl
Verbatim

really bad. all major Chinese models are being trained on Blackwell class chips

12 Nov 2026, 21:34 UTC · 2,813 likes · 175 repostsView post on X

The Chip Security Act and hardware telemetry

Because checkpoint forensics sees thread grids rather than addresses, Congress backed hardware telemetry. The Remote Access Security Act (RASA, H.R. 2683) passed the House 369–22 in January 2026 before stalling.

Alongside it sits the Chip Security Act (H.R. 3447 / S. 1705), clearing committee 42–0 in March 2026. CNAS and IAPS put ping telemetry under $1M upfront and $2.5M–$12.5M yearly.

US export controls, offshore compute gaps and Chip Security Act timeline (2025–2026)
Rule or billDateTechnical scope or sumOperational gap exposed
H200 narrow export rule15 Jan 2026<21,000 TPP, <6,500 GB/s, 50% capExcludes NVIDIA B200, B300, GB200 and GB300
RASA (H.R. 2683)12 Jan 2026 (369–22)Restricts remote US cloud trainingAwaits Senate passage; foreign clouds persist
Chip Security Act (H.R. 3447)26 Mar 2026 (42–0)Mandates tamper alerts & chip geolocationNot yet enacted into federal law
BIS subsidiary guidance31 May 2026Flags China-HQ affiliates on US cloudsExempts ongoing overseas training sites
DOJ smuggling indictments2025–Oct 2026$160M, $300M & $2.5B casesPhysical diversion bypasses paper manifests

Until Congress passes the Chip Security Act or tightens offshore rules, compute stays hidden in a regulatory warren. Even then, checkpoint forensics tracks silicon access rather than jurisdiction, shaping economics like DeepSeek’s API pricing.

folkfox vixen standing beside a brass compass over glowing cryptographic grid lines on dark rag paper

How labs erase the trail and what audits check next

Every side-channel audit on software defaults has a short shelf life. Publishing how SM stride marks weights gave engineering teams a checklist to wipe their quarry trail before future releases.

DePue weighed that disclosure risk with policy officials beforehand. Analysts @ManavhShah and @teortaxesTex predicted labs will anonymise starting weights to outfox future checkpoint forensics checks.

Will DePue@willdepue
Summary of X post

Notes that now the method is public any lab can erase many signatures, explaining he consulted advisers and officials before concluding publication remained worthwhile.

12 Nov 2026, 21:29 UTC · 462 likes · 13 repostsView post on X
Manav Shah@ManavhShah
Verbatim

@willdepue If the GPU leaves a fingerprint in open weights, every open-weights release is now an audit of where the chips came from. Expect labs to scramble their init before publishing.

12 Nov 2026, 23:41 UTCView post on X
Teortaxes@teortaxesTex
Summary of X post

Suggests observers will scrutinise DeepSeek’s V4-to-V4.1 transition for hardware anonymisation as Blackwell supply stays capped.

12 Nov 2026, 22:04 UTC · 5 likesView post on X
Will DePue@willdepue
Verbatim

There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon. Stay tuned.

12 Nov 2026, 21:30 UTC · 330 likes · 11 repostsView post on X

Four quick fixes that break checkpoint forensics

Erasing CUDA Philox geometry takes minutes without hurting model convergence:

  • CPU random generation before tensor upload
  • Element-indexed counter PRNG ignoring GPU blocks
  • Orthogonal basis rotation before step one
  • Non-CUDA engines such as PaddlePaddle seeding

Qwen3 leaves zero GPU trail via CPU seeding, while ERNIE-5.5 uses PaddlePaddle seed 2760. Dropping PyTorch CUDA defaults strips SM traces before checkpoint forensics runs.

Enterprise model audits after checkpoint forensics

Still, checkpoint forensics expands model provenance checks. ICLR 2026’s SeedPrints (GitHub code) and Proof-of-Learning verify lineage from initial weights.

For enterprise teams, folkfox AI consultancy and audit retainers separate verified checkpoint forensics provenance from claims whenever open weights cross borders.

Quick answers

Direct answers on checkpoint forensics, the six matched Chinese models, export rules and how labs hide PRNG layouts, each unpacked above.

What is checkpoint forensics in AI models?

Checkpoint forensics recovers surviving random initialisation layouts from published model weights to identify which GPU architecture initialised a model.

Which six models matched NVIDIA Blackwell layouts?

DeepSeek V4 Flash and MiniMax M3 matched 148-SM B200/B300 grids, while DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash matched 152-SM GB200/GB300 grids.

Does checkpoint forensics prove chips were smuggled into China?

No. Checkpoint forensics detects GPU thread counts rather than GPS location or legal ownership; chips could operate inside smuggled clusters or leased third-country clouds.

Why did critics question the $500 million cluster estimate?

Jonathan Sandhu and @teortaxesTex noted that dollar totals and 2,000–3,000 chip counts rely on assumed 35% utilisation and run duration rather than direct weight telemetry.

Could AMD MI300X GPUs mimic 148-SM or 152-SM signatures?

Yes. Partitioned MI300X GPUs can replicate those SM layouts; DePue ruled out AMD numerically on DeepSeek V4 Flash (363/377 NVIDIA matches), leaving peers partly open.

How can AI labs erase checkpoint forensics signatures?

Labs scrub checkpoint forensics traces by initialising weights on CPUs, using element-indexed counters, rotating bases before training or switching from PyTorch CUDA defaults.

Work with folkfox

Audit your AI stack before provenance becomes public

folkfox advises technical buyers on model provenance, open-weight supply chains, AI governance and engineering consultancy across European and global deployments.

Sources

Primary technical reports, PyTorch CUDA source files, model cards, US policy filings and DOJ court records checked 13 November 2026.

  1. Unemployment Labs: Checkpoint Forensics (12 November 2026)
  2. X root post: Will DePue (@willdepue, 12 November 2026)
  3. PyTorch ATen CUDA source: DistributionTemplates.h
  4. PyTorch documentation: Reproducibility and randomness
  5. NVIDIA cuRAND docs: Device API and Philox4x32-10
  6. PyTorch GitHub issue #196613: CUDA RNG SM grid layout
  7. Hugging Face: deepseek-ai/DeepSeek-V4-Flash
  8. Hugging Face: MiniMaxAI/MiniMax-M3
  9. arXiv 2606.13392: MiniMax Sparse Attention (June 2026)
  10. Hugging Face: inclusionAI/Ling-3.1-flash
  11. GitHub: inclusionAI/cuLA BENCHMARK_GB300.md (31 March 2026)
  12. US BIS: AI Diffusion rescission and controls update
  13. US BIS: China-headquartered subsidiary guidance (31 May 2026)
  14. Carnegie Endowment: Controlling cloud compute (May 2026)
  15. CNAS: Technology to secure the AI chip supply chain
  16. US DOJ: Over $300M AI server smuggling indictment
  17. IAPS: Location verification for AI chips
  18. ICLR 2026: SeedPrints model provenance poster
  19. GitHub: YnezT0311/SeedPrints provenance codebase
  20. arXiv 2103.05633: Proof-of-Learning training audit framework

Reporting: the folkfox newsroom, from the 20 primary sources listed above and archived X records, checked 13 November 2026. Illustrations are conceptual watercolour plates of the folkfox vixen made for this report. Nothing here is legal or financial advice.

The den

Where to?

Pricing

Choose a section. Enter opens it, Escape continues reading.

Cookie preferences

folkfox uses data the way we use strategy: only when it earns its place.