Six Chinese Open-Weight Models Carry a Blackwell Fingerprint, an Independent Weight Analysis Finds, but It Cannot Say Where the Chips Were
Will DePue of Unemployment Labs published an analysis on November 12 arguing that six models released since April 2026 by DeepSeek, Moonshot AI, MiniMax and Ant Group were initialized on Nvidia Blackwell-class GPUs, which need a U.S. licence for export to China. The method reads leftover starting values in the published weights and matches them to each chip's processor-core layout; earlier models from the same labs read as Hopper. DePue confirmed two matches on physical Blackwell hardware. He says the evidence shows access to the chips, not where they were located, how access was obtained, or whether export controls were violated.
Washington's arguments about whether Chinese labs train on Nvidia's newest chips have so far rested on officials' statements and anonymous sourcing. A new analysis claims to read the answer from the models themselves. Will DePue of Unemployment Labs published a report on November 12 saying six open-weight models released since April 2026 by DeepSeek, Moonshot AI, MiniMax and Ant Group carry numerical patterns that match Nvidia's Blackwell-class GPUs, which need a U.S. licence for export to China. It is technical evidence of access to the chips, but not of where they were or how they were obtained, as the author says himself.
What the report finds#
The six models are DeepSeek-V4-Flash and DeepSeek-V4-Pro, Moonshot's Kimi K3 and Kimi K3.1-Pro, MiniMax-M3 and Ant Group's Ling-3.1-flash. Earlier models from the same labs, such as DeepSeek-V3 and Kimi K2.5, read as Nvidia's older Hopper generation, the report says. The newest models it could read from Alibaba, Zhipu, Tencent, Xiaomi and StepFun also read as Hopper.
The method relies on a quirk of how PyTorch sets a model's random starting values on Nvidia GPUs: where each generated number lands depends on how many processor cores, or streaming multiprocessors, the chip has. Some starting values survive training almost untouched, so the report searches for the seed and chip layout that reproduce them, then tests that combination on weights held out of the search. For DeepSeek-V4-Flash and Kimi K3, DePue says he reproduced the preserved values on physical Blackwell chips. He reports that the method agreed with all 24 cases where a developer had publicly named a model's hardware.
The report notes that U.S. officials and press reports had already alleged DeepSeek and Moonshot were using Blackwell chips, without publishing technical evidence. Unite.AI's write-up of the report describes the same findings and limits.
What it does not show#
The fingerprint identifies the hardware that created a model's starting weights. DePue says this supports, without proving, that Blackwell also did the training. The only alternative he identified is specially configured AMD hardware, also export-controlled, which he excludes for DeepSeek-V4-Flash and considers very unlikely for the other five. Crucially, Blackwell systems can be rented in data centres outside China, which Reuters reported this week Chinese labs are rushing to do.
Limits#
- This is one researcher's analysis. The reproduction scripts are available to reviewers on request, and we have not seen an independent replication.
- The method cannot establish chip ownership, location or acquisition, and so cannot show smuggling or an export-control violation, as the report states.
- The 24-case agreement is a consistency check on related, publicly documented models, not a measured accuracy rate.
- DeepSeek, Moonshot AI, MiniMax, Ant Group and Nvidia have not responded in the material we read.