DeepSeek AI Model Files Point to Training on Nvidia Blackwell Chips

Numerical traces in DeepSeek’s public AI model files point to access to Nvidia’s Blackwell processors, according to a November 12 analysis by researcher Will DePue. His method reads patterns left when a model’s starting values were created, turning downloadable files into clues about the computing hardware behind them.

DePue’s self-published report at Unemployment Labs covers six models from DeepSeek, Moonshot AI, MiniMax and Ant Group. Its strongest Nvidia-specific result concerns DeepSeek-V4-Flash.

How Starting Values Reveal Hardware

A model’s weights are the adjustable numbers that determine how it processes information and produces answers. Before training can change them, software supplies their starting values, a process called initialization. Some weights, including values reserved for symbols never encountered during training, can retain those originals. Others preserve a weaker statistical resemblance.

In the software DePue examined, a number called a seed starts a generator of seemingly random values. The same seed and software settings reproduce the same sequences. PyTorch, a framework for developing AI models, spreads that work across parallel software workers called threads on a graphics processor, or GPU. Its initialization implementation uses the GPU’s number of processing units, called streaming multiprocessors, and their thread capacity to determine how the work is divided.

Changing the number of workers changes where their numbers land. In DePue’s simplified example, one worker fills positions 1, 49 and 97 when 48 workers share the job. With 40 workers, the same sequence lands at positions 1, 41 and 81. Recovering both the seed and the arrangement can therefore identify the hardware layout that generated surviving values.

DeepSeek Supplies the Clearest Physical Match

DePue reports recovering the same seed in DeepSeek-V3 and V4-Flash, then running unmodified PyTorch on two physical Nvidia chips. An older Hopper H100 reproduced V3’s preserved values. A Blackwell B300 reproduced 445 of V4-Flash’s 469 selected weight rows bit for bit. Neither chip reproduced the other model’s preserved rows, and changing the seed by one broke the matches.

For Moonshot’s Kimi K3, he reports that a rented GB200 reproduced all 13 preserved rows tested.

He searched roughly six trillion seed-and-layout combinations, then tested recovered settings on weights held out of the search. His statistical model put the strongest accidental match near 8 sigma, or standard deviations above random variation; the strongest result for each of the six models ranged from 49 to 1,385 sigma.

Six Models, With Different Levels of Certainty

The report finds older Hopper-compatible patterns in earlier models from the same four developers. Its six main Blackwell-compatible results fall into two groups: 148 streaming multiprocessors for B200/B300, and 152 for GB200/GB300. The exact chip within each pair remains unresolved.

AMD’s MI300X GPU can also produce those arrangements. DePue tested a physical MI300X with only 148 or 152 of its usual 304 compute units active and reproduced the layouts.

A separate arithmetic difference between Nvidia and AMD helped distinguish V4-Flash. Among 377 preserved values for which the vendors produced different answers, 363 matched Nvidia and none matched AMD, DePue reports. That distinction remained unresolved for the other five models. He considers AMD unlikely for them because reproducing the pattern would require an unusual, consistent configuration across the initializing cluster.

What the Fingerprints Can Follow

An initialization fingerprint can survive a transfer to different training hardware or be inherited by a later model. Moonshot’s Kimi K3.1-Pro is post-trained from K3, so six released models need not represent six independent training runs.

The technique needs public weights that retain a recognizable GPU-generated pattern. DePue reports recovering no such signature in DeepSeek-V4.1-Flash.

From model size and training-data volume, DePue estimates that V4-Flash would require 1,000 to 1,500 Blackwell GPUs for six to eight weeks. That calculation assumes the initializing hardware also trained the model, operating at 35% of rated throughput. His broader uncertainty range is 250 to 4,500 GPUs.

DePue offers the reproduction scripts, row fixtures and test outputs to reviewers on request.

Earlier Access Allegations and Licensing Rules

In December 2025, The Information reported that DeepSeek acquired about 2,000 Blackwell chips.

On July 22, 2026, White House science and technology policy director Michael Kratsios alleged that Moonshot had acquired GB300-equipped servers and accessed GB300s in Thailand, likely to train its models.

In its filing for the quarter ended July 26, 2026, Nvidia lists B200 and GB200 NVL72 among Blackwell products requiring export licenses for China.

U.S. licensing requirements can also apply to covered advanced-computing items exported to companies outside China. Commerce Department guidance dated May 31, 2026 covers recipients headquartered in China or whose ultimate parent company, at the top of their corporate group, is headquartered there.

The model files leave the inferred hardware’s physical location, ownership and access arrangements unknown. Whether any export controls were violated remains unresolved.