Open Chinese AI models allegedly show traces of training on export-restricted Blackwell GPUs

Independent researchers claim to have found fingerprints of Blackwell GPU training in the weights of several open Chinese AI models, potentially implicating export controls.
According to independent researchers, open model weights may retain a fingerprint of the GPU on which they were first generated. Initial weight values depend on the seed and the specific chip, and part of that pattern can survive training.
The analysis reportedly examined about six trillion combinations of chips and seeds and found matches across six models from four Chinese labs:
- DeepSeek-V4-Flash and MiniMax-M3 were allegedly trained on B200 or B300 GPUs.
- DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash were allegedly trained on GB200 or GB300 GPUs.
Earlier models from the same labs reportedly appear to have been trained on Hopper.
The author also claims that Kimi K3 and DeepSeek-V4-Pro may have used 2,000 to 3,000 Blackwell GPUs each. A cluster of that scale costs about $200 million.
The author says the findings were checked by researchers from OpenAI and Anthropic. If confirmed, Chinese labs could face serious sanctions.
Why it matters
AnalysisIf true, hardware provenance may be detectable in model weights, making export-control enforcement more feasible.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
DeepSeek (Chinese Breakthrough in Open AI)
A series of groundbreaking open models from the Chinese lab DeepSeek (V3, R1). Demonstrates top-tier coding and reasoning capabilities at a cost 10–20 times lower than Western commercial counterparts.
GPT Image Generative Models Prompting Guide
OpenAI's gpt-image generative models are designed to produce production-quality visuals and fully controllable creative workflows. They are suitable for both professional design ta...
DeepSeek open-sources part of its chat interface
DeepSeek has released the source code for a portion of its chat UI, including web search, deep thinking, and conversation history features.
Claude Sonnet 4.6 vs GPT-5.5 vs Gemini 3.1 Pro: 2026 model comparison
Every few weeks, a new model comes out, and every time the same questions arise: should I switch? what changed what to pay for?