





A folkfox special report.
Checkpoint forensics: the six hidden Blackwell clues inside open weights
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.







01
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
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.
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. [...]
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
| Tier | SMs | Threads | Models |
|---|---|---|---|
| NVIDIA Hopper (H100 / H800 / H200) | 132 SMs | 270,336 threads | DeepSeek-V3, MiniMax-M2.7, Ling-3.0-flash |
| NVIDIA Blackwell (B200 / B300) | 148 SMs | 303,104 threads | DeepSeek V4 Flash, MiniMax M3 |
| Grace Blackwell (GB200 / GB300) | 152 SMs | 311,296 threads | V4-Pro, Kimi K3, K3.1-Pro, Ling-3.1-flash |
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.
@willdepue @aidenybai GPUs use a Philox PRNG.. so the layout of SM thread/block/warp leaves a geometric signature matching the kernel layout?
@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.






02
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
| Model | Best Z-score | Matched grid |
|---|---|---|
| DeepSeek-V4-Flash | 1385 sigma | 148 SM (B200/B300) |
| MiniMax-M3 | 323 sigma | 148 SM (B200/B300) |
| Kimi K3 | 305 sigma | 152 SM (GB200/GB300) |
| Kimi K3.1-Pro | 305 sigma | 152 SM (GB200/GB300) |
| Kimi K2.7-Code | 246 sigma | 80/20 132+148 blend |
| Ling-3.1-flash | 244 sigma | 152 SM (GB200/GB300) |
| DeepSeek-V4-Pro | 49 sigma | 152 SM (GB200/GB300) |
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
[...] This is that search on DeepSeek-V4-Flash. Every processor count gives noise until 148 (Blackwell), where the weights match.
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.
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.
| Model & lab | Release window | Matched grid & sigma | Corroboration & caveat |
|---|---|---|---|
| DeepSeek V4 Flash (DeepSeek) | Apr 2026 (284B / 13B) | 148 SM (B200/B300), 1,385 sigma | 445/469 exact rows on B300; 363/377 NVIDIA vs 0 AMD |
| DeepSeek-V4-Pro (DeepSeek) | Apr 2026 | 152 SM (GB200/GB300), 49 sigma | Weakest flagship score; AMD edge case unresolved |
| Kimi K3 & K3.1-Pro (Moonshot) | Jul & Nov 2026 | 152 SM (GB200/GB300), 305 sigma | 13/13 Kimi K3 rows on GB200 at r = 0.999995 |
| Kimi K2.7-Code (Moonshot) | Jun 2026 | 80/20 Hopper+148 SM, 246 sigma | Blended Hopper-to-Blackwell transition release |
| MiniMax M3 (MiniMax) | Jun 2026 (428B / 23B) | 148 SM (B200/B300), 323 sigma | M2.7 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 sigma | Ling-3.0-flash matched 132 SM; cuLA log dated 31 Mar |
| DeepSeek-V4.1-Flash & Qwen3 | Control suite | Empty GPU fingerprint | Non-CUDA or CPU init proves neither NVIDIA nor Ascend |






03
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
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
Separates releasing public Blackwell inference kernels from financing billion-dollar training runs on restricted silicon.
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.
Separates sound Philox engineering from procurement claims: init leaves location unproven, MI300X/GB300 overlap remains open, and $500M+ assumes 35% utilisation.
Notes 2,000–3,000 Blackwells imply sub-10MW clusters, while agreeing checkpoint forensics confirms active Blackwell stacks.
folkfox chart
Model scale across the 2026 Hopper-to-Blackwell shift (billions)
| Model | Total parameters | Active per token | Matched layout |
|---|---|---|---|
| Ling-3.1-flash | 560B | 25B | 152 SM · GB200/GB300 |
| MiniMax M3 | 428B | 23B | 148 SM · B200/B300 |
| DeepSeek V4 Flash | 284B | 13B | 148 SM · B200/B300 |
| MiniMax-M2.7 | 229B | 10B | 132 SM · Hopper |
| Ling-3.0-flash | 124B | 5.1B | 132 SM · Hopper |
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.
| Dimension | Observed evidence | Unresolved gap | Confidence level |
|---|---|---|---|
| Initialisation match | 445/469 V4-Flash rows exact on B300; 13/13 Kimi K3 on GB200 | Public repo pending; scripts currently by request | Very high (V4-Flash & Kimi K3) |
| NVIDIA vs AMD MI300X | 363/377 NVIDIA cuRAND vs 0 AMD matches on V4-Flash | Value-level AMD test unconfirmed for all 5 peers | High (V4-Flash); moderate elsewhere |
| Full training run | Residual correlation r = 0.066 (57 sigma) in trained layers | Cross-chip transfer after init unruled out empirically | Strong inference, not proof |
| Cluster count & spend | Modelled ~2k–3k Blackwells & $500M+ across labs | Assumes 35% MFU utilisation and training cadence | Modelled estimate only |
| Physical location | Six models carry 148/152-SM Blackwell signatures | Weights cannot tell smuggled chips from foreign clouds | Zero direct visibility |






04
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.
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.
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.
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.
really bad. all major Chinese models are being trained on Blackwell class chips
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.
| Rule or bill | Date | Technical scope or sum | Operational gap exposed |
|---|---|---|---|
| H200 narrow export rule | 15 Jan 2026 | <21,000 TPP, <6,500 GB/s, 50% cap | Excludes NVIDIA B200, B300, GB200 and GB300 |
| RASA (H.R. 2683) | 12 Jan 2026 (369–22) | Restricts remote US cloud training | Awaits Senate passage; foreign clouds persist |
| Chip Security Act (H.R. 3447) | 26 Mar 2026 (42–0) | Mandates tamper alerts & chip geolocation | Not yet enacted into federal law |
| BIS subsidiary guidance | 31 May 2026 | Flags China-HQ affiliates on US clouds | Exempts ongoing overseas training sites |
| DOJ smuggling indictments | 2025–Oct 2026 | $160M, $300M & $2.5B cases | Physical 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.






05
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.
Notes that now the method is public any lab can erase many signatures, explaining he consulted advisers and officials before concluding publication remained worthwhile.
@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.
Suggests observers will scrutinise DeepSeek’s V4-to-V4.1 transition for hardware anonymisation as Blackwell supply stays capped.
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.
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.
Keep reading
Read more on this topic
Nvidia and the China spy allegation: what is known?
US intelligence officials reportedly believed an Nvidia business contact was passing information to China. Nvidia disputes that sensitive information was shared.
Read the pieceModel infrastructureDeepSeek V4 Just Released a Model Named After Its Own Expiry Date
DeepSeek V4.1 Flash hits 400 tokens a second and expires on 10 September. Five honest checks for buyers before the beta becomes the default API route.
Read the pieceOpen weightsOpen weight models meet a 3,800-GPU reality. Mistral’s preview exposes the infrastructure underneath
Open weight models can loosen API dependence, not the supply chain. Mistral Large 4’s preview puts GPUs, memory and European control under practical scrutiny.
Read the pieceWork 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.
Want folkfox in your Google results? Set folkfox as a preferred source.
Sources
Primary technical reports, PyTorch CUDA source files, model cards, US policy filings and DOJ court records checked 13 November 2026.
- Unemployment Labs: Checkpoint Forensics (12 November 2026)
- X root post: Will DePue (@willdepue, 12 November 2026)
- PyTorch ATen CUDA source: DistributionTemplates.h
- PyTorch documentation: Reproducibility and randomness
- NVIDIA cuRAND docs: Device API and Philox4x32-10
- PyTorch GitHub issue #196613: CUDA RNG SM grid layout
- Hugging Face: deepseek-ai/DeepSeek-V4-Flash
- Hugging Face: MiniMaxAI/MiniMax-M3
- arXiv 2606.13392: MiniMax Sparse Attention (June 2026)
- Hugging Face: inclusionAI/Ling-3.1-flash
- GitHub: inclusionAI/cuLA BENCHMARK_GB300.md (31 March 2026)
- US BIS: AI Diffusion rescission and controls update
- US BIS: China-headquartered subsidiary guidance (31 May 2026)
- Carnegie Endowment: Controlling cloud compute (May 2026)
- CNAS: Technology to secure the AI chip supply chain
- US DOJ: Over $300M AI server smuggling indictment
- IAPS: Location verification for AI chips
- ICLR 2026: SeedPrints model provenance poster
- GitHub: YnezT0311/SeedPrints provenance codebase
- 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.

