The AI Read is on a break. The last edition was November 15, 2026. Everything here stays open to read.
The AI Read
← AI News

Illustration, AI-generated from the story text. Tap a panel to read the story through it.

Morning Brief · November 13th 2026

Six open-weight models from DeepSeek, Moonshot, MiniMax and Ant carry a Blackwell fingerprint, a forensic analysis finds

Will DePue of Unemployment Labs published a method on Thursday for reading which kind of GPU produced a model's starting weights. It works because the random-number layout depends on how many streaming multiprocessors the chip has: 132 on Hopper, 148 on B200 and B300, 152 on GB200 and GB300. Six Chinese models read as Blackwell. DeepSeek-V4-Flash and MiniMax-M3 match B200 or B300; DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ant's Ling-3.1-flash match GB200 or GB300. DePue reproduced the V4-Flash result on a physical B300 and the Kimi K3 result on a rented GB200. The newest models from Alibaba, Zhipu, Tencent, Xiaomi and StepFun read as Hopper.

The report is careful about its limits. It shows where weights were initialized, which is not necessarily where the model was trained, and it cannot say where the chips sat or how they were obtained; Blackwell capacity can be rented legally outside China. No lab and no one at Nvidia has responded, and nobody has replicated it. If it holds, enforcement gets a test that runs on the published weights. SourcesAC

Sources

This story ran in the Morning Brief of November 13th 2026.