Annanim News
Now
CybersecurityMore stories: Apache Roller, F5 flaws, Rockwell and AGILOX, and quantum researchCybersecurityUkraine and the FBI arrest four suspects in a ransomware gang that collected more than $40 millionQuantum & precision measurementNIST shows the first complete quantum position verification: quantum physics confirms where a person isAutonomous vehiclesTesla tightens driver monitoring in its European driving system: hands on the wheel above 130 km/hCybersecurityPalo Alto Networks to acquire Applied Compute for $4.5 billion in stockModels & researchGoogle launches Gemini 4 Neon, a fast model for code, office work and agentsChinaChina Watch: Tencent raises its computing infrastructure spending sixfoldCybersecurityDecade-old NGINX flaw lets attackers run code on HTTP/2 servers without logging inRegulation & businessAnthropic prices its IPO at $83 a share, above the range, and starts trading on Nasdaq todayCybersecurityMore stories: cracking BitLocker PINs, Russian hackers and Nvidia's guaranteesSpace, satellites & commsVoyager 1 becomes the first human-made object one light-day from EarthCybersecurityIsrael's Vero Security raises $11 million to secure enterprise AI agentsAI agentsGoogle turns its Spark agent into Tasks inside Gemini and makes it free

Originally published:

Study: DeepSeek, Kimi and other Chinese models were trained on export-banned Blackwell chips

Research group Unemployment Labs, led by former OpenAI researcher Will DePue, analyzed the weights of Chinese open models and found fingerprints of training on Blackwell chips, which may not be exported to China. The researchers stress the finding does not show where the chips were or that export rules were broken.

Study: DeepSeek, Kimi and other Chinese models were trained on export-banned Blackwell chips
Image: TechCrunch (opens the source site in a new tab)

The method exploits the weight initialization step. Random initialization is split across the chip's streaming multiprocessors (SMs), and each chip family has a different count: 132 on Hopper, 148 on Blackwell B and 152 on Blackwell GB. The researchers recreated the process with the same chips and software and found the resulting weight distribution precisely matched that of the Chinese models. When they changed the SM count, the match disappeared. This let them identify the chip family for each model, as well as an approximate number of cards.

Models examined include DeepSeek V4, Moonshot's Kimi K3.1, MiniMax-M3 and Ant Group's Ling-3.1-Flash. The method works only on open-weight models and does not show whether training ran entirely on Nvidia chips. The chips may have been in a data center outside China, for example in Malaysia; according to Tom's Hardware, it is currently not illegal for Chinese companies to rent Blackwell hardware abroad.

The research comes a day after Jensen Huang resigned as Nvidia CEO amid a scandal tied to a company accused of smuggling chips into China. Epoch AI estimates that about 2 million advanced chips entered China in 2024 and 2025 through schemes that hid their final destination.

Sources

Topics

Install the app

In Safari: tap the Share button, then “Add to Home Screen”.