Open-Weight Models

Checkpoint forensics ties six Chinese open-weight models to Blackwell GPUs

Unemployment Labs matched initialization traces in released weights to GPU thread counts. Six DeepSeek, Kimi, MiniMax and Ling models match export-controlled NVIDIA Blackwell chips.

Published
November 13, 2026
Read
6 min
Published by
Samir Sengupta
Checkpoint forensics matches six Chinese open-weight models to NVIDIA Blackwell GPU thread counts
Note 143 / 147Daily note · Written from 1 source

How this was madeDrafted with Claude, Anthropic’s model, from 1 cited source, then checked against it in a second editing pass. Editorial policy

the short version

  • For sufficiently large initializations on the NVIDIA chips examined, PyTorch's cuRAND-backed Philox initialization uses 2,048 threads per SM. The layout of starting values therefore encodes the SM count of the chip that created the weights.
  • Six models released since April 2026, including DeepSeek-V4-Pro and Kimi K3, match Blackwell layouts at 49 to 1,385 sigma. The strongest chance match across the search would be expected to reach about 8 sigma.
  • The method agreed with all 24 models whose developers have disclosed their hardware. The fingerprint shows how the weights were created. It supports, but does not prove, Blackwell training, and it says nothing about chip location or export-control violations.
  • The GPU counts derived from these runs are soft. The report has already revised them once, and they depend on run lengths and utilization figures that no lab has published.

Checkpoint forensics on released open weights finds that six models from DeepSeek, Moonshot AI, MiniMax and Ant Group, all released since April 2026, contain initialization patterns that match NVIDIA Blackwell GPUs. Unemployment Labs published the analysis on November 12, 2026. The method recovers traces of each model's starting values from the published weights and searches for the seed and GPU thread count that reproduce them. The six matches score between 49 and 1,385 sigma, against about 8 sigma for the strongest match expected by chance. On 24 models whose developers have disclosed their hardware, the method agreed every time.

Earlier models from the same four labs read as Hopper. DeepSeek switched between DeepSeek-V3 (December 2024) and V4 (April 2026). Ant Group switched between Ling-3.0-flash (August 2026) and Ling-3.1-flash (October 2026). Exporting Blackwell to China requires a U.S. license, and the report notes that none has been reported. It also says the method cannot establish where the chips were, how access was obtained, or whether export controls were violated.

For engineers who publish weights, there is a second point. The arrangement of preserved initial values in a checkpoint carries a hardware signature that can be recovered from the published weights alone. The report draws on no leaks or insider accounts.

How checkpoint forensics reads the GPU

PyTorch fills initial weights using Philox, a deterministic pseudorandom generator, through NVIDIA's cuRAND library. The work is split across GPU threads. On the NVIDIA chips examined, a sufficiently large initialization runs the maximum of 2,048 threads per streaming multiprocessor (SM), and each thread writes its own sequence into strided positions. A chip with more SMs uses more threads, so the same seed puts the same numbers in different places. In the report's toy example, going from 48 threads to 40 leaves only 6 percent of the displayed values matching the original arrangement.

  • Hopper H100 SXM, H200, H800: 132 SMs, 270,336 threads
  • Blackwell B200/B300: 148 SMs, 303,104 threads
  • Blackwell GB200/GB300: 152 SMs, 311,296 threads

Reproducing a weight layout requires both the seed and the thread count to be correct. The search covered roughly 6 trillion seed-and-layout combinations. Part of the signal can come from weights that training leaves untouched, such as those reserved for words or symbols never encountered during training, which keep their starting values almost exactly. The available excerpt breaks off partway through explaining how these preserved values are compared, so the full recovery procedure is not described here. The result is a hardware group, not a specific chip: a Hopper match cannot tell an H800 from an H100 or H200.

One commenter on Hacker News questioned how robust the fingerprinting is without a PyTorch bug, since parallel Philox generation is a well-researched area. The report's explanation does not rest on a weakness in Philox. It rests on how PyTorch, through cuRAND, divides Philox output among threads, which makes the thread count determine where each value lands. Another commenter pointed out that the report links no code. The excerpt does not say whether the forensic code will be released, and the report states that its token-count estimation method has not yet been published.

Six models from four labs read as Blackwell

  • DeepSeek-V4-Flash (Apr 2026): B200/B300. DeepSeek-V4-Pro (Apr 2026): GB200/GB300. DeepSeek-V3 (Dec 2024) read as Hopper.
  • Kimi K3 (Jul 2026) and Kimi K3.1-Pro (Nov 2026): GB200/GB300. Kimi K2 (Jul 2025) and K2.5 (Jan 2026) read as Hopper. K3.1-Pro starts from the same seed as K3 and may be a checkpoint of K3 with further post-training.
  • MiniMax-M3 (Jun 2026): B200/B300. MiniMax-M2.7 (Apr 2026) read as Hopper.
  • Ling-3.1-flash (Oct 2026): GB200/GB300. Ling-3.0-flash (Aug 2026) read as Hopper.

Kimi K2.7-Code, posted six weeks before K3, falls in between. It is built on the Hopper-era weights but contains a small part that matches the B200 or B300. The newest models the authors could read from Alibaba (Qwen), Zhipu (GLM), Tencent, Xiaomi and StepFun all read as Hopper. Meituan says LongCat-2.0 was trained entirely on "AI ASIC superpods," and the report found no NVIDIA fingerprint in the weights it has released.

The method matched all 24 disclosed models

For 24 models whose developers have publicly named the hardware used, the method's reading agreed with the disclosure in every case. NVIDIA's Nemotron-3 Ultra, whose technical report names the GB200 for one phase of training, reads as 152 SMs, the same as Kimi K3. For DeepSeek-V4-Flash and Kimi K3, the authors also confirmed the match on physical Blackwell chips. The weakest individual test reported reaches 20 sigma, against the roughly 8 sigma expected from the best chance match in the search.

The validation shows the readings are accurate where the answer is known. The excerpt does not list the 24 models or say which hardware groups they span. It does not report how many models yielded no readable fingerprint, and it does not give a false-negative rate.

Initialization hardware is not training hardware


The fingerprint identifies how the weights were created and supports, without proving, the inference that Blackwell also performed the training.
Will DePue, Unemployment Labs

The report sets out several limits. A Hopper reading cannot tell an H800 from an H100 SXM or H200, so the H800 labels in its table are lab disclosures, not findings of the method. The only alternative hardware the authors identified is specially configured AMD hardware, which is also export-restricted. They excluded it for DeepSeek-V4-Flash and consider it very unlikely for the other five models.

Commenters on Hacker News argued that training on Blackwell capacity rented outside China, for example in Singapore or Malaysia, can be legal. The report does not address that legal question. It says only that the method does not establish where the chips were located or how access was obtained.

Cluster estimates of 3,000 to 4,500 GPUs are soft

The report assumes training stayed on the hardware that initialized each model. On that basis, its best estimates are about 4,500 Blackwell GPUs for Kimi K3 (range 1,000 to 9,500) and 3,000 for DeepSeek-V4-Pro (range 750 to 6,500), each over about 70 days. Across the five runs it estimates, the total is about 12,000 Blackwell GPUs (range 7,000 to 22,000), or about 30,000 H100 equivalents. These figures assume about 20 percent utilization, or 25 percent for DeepSeek. At roughly $70,000 per GPU, a figure derived from a Wall Street Journal report on Blackwell chips installed in Malaysia for ByteDance, the largest cluster would cost $200 million to $300 million and the five together more than half a billion dollars.

The report revised these numbers on publication day. Its first version gave 2,000 to 3,000 GPUs for the largest runs and $150 million to $200 million for the largest cluster, because it had assumed NVIDIA benchmark efficiency. Run length matters most, since halving it doubles the GPU count, and no lab has stated its run lengths. For comparison, Mistral says it trained Mistral Large 4 on 3,800 Grace Blackwell GPUs, and Microsoft says it trained MAI-Base-2 on 16,000 in 38 days. At Microsoft's pace, the two largest runs here would have needed about 8,000 to 9,000 GPUs.

The figures cover only each model's final training run, not experiments, earlier runs or serving. Any capacity comparison built on them should use the ranges rather than the point estimates.

What initialization layout reveals in released checkpoints

The mechanism is not specific to Chinese labs. Readings from the 24 disclosed models, including NVIDIA's Nemotron-3 Ultra, show it works on other developers' checkpoints. A model initialized through PyTorch's cuRAND path on these NVIDIA configurations gets a layout tied to the chip's SM count. Weights that training leaves untouched can preserve that layout in the released checkpoint. The shared seed between Kimi K3 and K3.1-Pro suggests the same signal can also point to lineage between releases.

The excerpt does not say whether reinitializing untouched weights, initializing on CPU, or changing the launch configuration would remove the signal. Teams that need to keep their training hardware private should not assume their published checkpoints are neutral.

Questions this raises

How does checkpoint forensics identify the GPU used to initialize a model?

PyTorch generates initial weights with the Philox generator through cuRAND and splits the work across GPU threads. A chip with more streaming multiprocessors runs more threads, so the same seed places values in different positions. Searching for the seed and thread count that reproduce preserved initial values reveals the hardware group.

Which Chinese open-weight models read as NVIDIA Blackwell?

DeepSeek-V4-Flash and MiniMax-M3 read as B200/B300. DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash read as GB200/GB300. Earlier models from the same labs read as Hopper.

Does the report prove export controls were violated?

No. The report says the method cannot establish where the chips were, how access was obtained, or whether export controls were violated. It identifies the initialization hardware, which supports but does not prove that Blackwell also performed the training.

Spotted an error? Email samir.s.sengupta@gmail.com and I will correct it and date the change at the top of the note. How these notes are researched, drafted and checked is in the editorial policy.

More notes

Building something on this?

I ship production LLM, RAG and agentic systems for a living - the infrastructure behind the things these notes are about. Open to roles, contract work and research collaboration.