AIToday
Large Language ModelsAI Regulation & PolicyAI Business & IndustryTechCrunch AIPublished: Nov 14, 2026, 04:00 JST

Blackwell fingerprints found in DeepSeek V4 and Kimi K3

Send on LINE
Blackwell fingerprints found in DeepSeek V4 and Kimi K3

3 Key Points

  1. What happened

    Will DePue of Unemployment Labs examined the published weights of open models from DeepSeek, Moonshot AI, MiniMax and Ant Group and found compelling evidence that Moonshot's Kimi K3 and DeepSeek's V4 models carry fingerprints of training on export-banned Blackwell chips.

  2. Why it matters

    The finding suggests the most closely watched open models were trained on chips that US rules bar from being shipped there, though the chips could have sat in an offshore data center, so a violation is not established.

WHO IT HITSUS export-control officials and compliance teams at chip designers and cloud providers, who now face pressure to track where controlled accelerators are actually used rather than only who owns them, may see this research sharpen that debate.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The finding lands in the middle of a long-running argument over whether export bans actually keep the most powerful chips out of Chinese labs' hands. The body notes that Epoch AI estimates about two million advanced chips entered China in 2024 and 2025 through schemes designed to hide their final destination, and that 275,000 Nvidia AI accelerators were imported to Malaysia in 2024 and 2025, a route that could place banned hardware just offshore.

The story also ties the chip question to corporate and political scrutiny. Nvidia CEO Jensen Huang left the company this week after admitting to an affair with Alice Huang, an executive of Megaspeed, a company implicated in smuggling Nvidia chips into China; Megaspeed's subsidiary Speedmatrix has been under investigation by the US and Singaporean governments. In May, short-seller Culper published a report arguing Nvidia had a "China problem" because a significant share of its revenue was driven by Chinese demand, and the body notes that some 600,000 H200 chips are reportedly stockpiled in Hong Kong awaiting shipment into China.

How the US government handles the matter is also in play: Secretary of Commerce Howard Lutnick, who oversees the agencies that license chip exports, could reportedly lose his job after reports that his investigators failed to uncover suspicious chip buyers and were too cozy with Nvidia officials. Against that backdrop, the weight analysis gives export-control enforcers a concrete technical signal to weigh, even as the body cautions that an offshore data center would explain the fingerprints without any violation.

FAQ
How did the researcher reach this conclusion?
Will DePue analyzed the weights of openly released large language models and found that the weights still reflect the configurations of the chips used to train them, with fingerprints pointing to Blackwell.
Does this prove export controls were broken?
No. The models may not have been trained entirely on Nvidia chips, and the chips could have been in an offshore data center, which has been a popular route for Chinese firms.
Which companies' models were examined?
Models released openly by DeepSeek, Moonshot AI, MiniMax and Ant Group were analyzed, with Kimi K3 and DeepSeek V4 named as showing the fingerprints.
Send on LINE

Also reported by THE DECODER

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleAmazon Bedrock adds per-user cost attribution in CUR 2.0