AI News

Trending story

Unemployment Labs Releases Analysis of Blackwell Initialization in Chinese Open-Source Models

7 reports1 discussions8 story updatesUpdated Nov 14 17:54

Understand this story

AI overview

On November 12, 2026, Unemployment Labs released a forensic report stating that the public weights of six Chinese open-source models—DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3, and Ant Group's Ling-3.1-flash—retain initialization layouts consistent with NVIDIA Blackwell chips (148 or 152 SMs), while earlier models in the same series correspond to Hopper. The report says the method is based on the fact that when PyTorch uses Philox to generate initial weights, the thread count varies with the number of SMs, and it reproduced the retained weights of DeepSeek-V3, V4-Flash, and Kimi K3 on physical H100, B300, and rented GB200 hardware. The report emphasizes that this evidence only indicates the initialization hardware and cannot determine the location of the chips, how they were obtained, or whether export controls were violated. The next day, reports said former OpenAI engineer Will DePue stated that after searching more than six trillion combinations of chip layouts and random seeds, the six models matched Blackwell, with four of them corresponding to GB200 and GB300, and estimated that Kimi K3 and DeepSeek-V4-Pro each used 2,000 to 3,000 Blackwell GPUs.

AI-generated from public reports

Latest development · Report

Researchers found traces of Nvidia GPU training in the weights of Kimi K3 and DeepSeek V4

Developments

1 confirmed developments

  1. Unemployment Labs releases analysis of Blackwell initialization patterns in Chinese open-source models

    Source:indigo

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 14

  1. t3n

    Researchers found traces of Nvidia GPU training in the weights of Kimi K3 and DeepSeek V4

    Will DePue, head of Unemployment Labs, said after analyzing Chinese model weights that he found traces of the Kimi K3 and DeepSeek V4 models being trained on Nvidia GPUs, chips that were originally subject to export controls. The report also noted that this does not show the models were trained only on Nvidia chips, nor does it directly establish a violation of export law; the chips could also be located in overseas data centers. Epoch AI estimates that from 2024 to 2025, about 2 million high-end chips entered China through circuitous channels.

  2. BaoyuRelated discussion

    Baoyu joined this discussion

    This post contributes to heat; it is not a formal report or confirmed fact.

  3. TechCrunch AI

    Study says Chinese open-source models such as DeepSeek and Kimi were trained on banned Nvidia Blackwell chips

    Will DePue, head of the nonprofit research organization Unemployment Labs, analyzed the weights of open-source models from DeepSeek, Moonshot AI, MiniMax, and Ant Group, and found that models such as Kimi K3 and DeepSeek V4 carry fingerprints of having been trained on Blackwell chips, which the United States has banned from export to China. The article also mentions that Nvidia CEO Jensen Huang stepped down this week due to his relationship with Megaspeed executive Alice Huang, and that Epoch AI estimates that about 2 million advanced chips entered China through covert channels between 2024 and 2025.

  4. cnBeta

    Study claims six model weights, including DeepSeek and Moonshot, carry NVIDIA Blackwell chip fingerprints

    Will DePue, a researcher at Unemployment Labs, released a technical investigation report on November 12, claiming that digital features matching NVIDIA's Blackwell architecture appear in six model weights released in recent years by DeepSeek, Moonshot, MiniMax and Ant Group, while earlier models mainly exhibited Hopper architecture features. The study exploits differences in the initial weights generated by PyTorch's Philox pseudorandom number algorithm on GPUs with different numbers of SMs to reverse-identify the random seed and hardware layout, and reproduced the corresponding weights of DeepSeek V3 and V4-Flash on H100 and B300. Based on this, the report estimates that some training clusters require hundreds to thousands of Blackwell GPUs, but the authors point out that the fingerprint only reflects the initialization hardware and cannot prove the chip's location or any violation of export controls.

Nov 13

  1. The Decoder

    Former OpenAI researcher says Nvidia Blackwell chip fingerprints detected in six Chinese AI models

    Former OpenAI researcher Will DePue says that through forensic analysis of published model weights, six models—DeepSeek-V4-Flash, MiniMax-M3, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, and Ant Ling-3.1-flash—carry initialization signatures of Nvidia Blackwell, with the first two corresponding to 148-core B200/B300 and the rest to 152-core GB200/GB300. He says the method identified all 24 control models with known hardware correctly, and the verification code has been published on GitHub. The report estimates that these five training runs together required about 12,000 Blackwell GPUs, costing more than $500 million, but the fingerprint can only indicate what hardware initialized the weights and cannot prove smuggling or a violation of U.S. export controls.

  2. indigoSelected

    Former OpenAI Engineer Says DeepSeek and Other Chinese Labs Trained Models on Blackwell Chips

    Former OpenAI engineer will depue posted that, through model weight initialization fingerprint analysis, he found that the latest Chinese models from DeepSeek, Moonshot, MiniMax and Ant Group were initialized and trained on NVIDIA Blackwell chips. He searched more than six trillion chip layout and random seed combinations, and said that six models including DeepSeek-V4-Pro, Kimi K3 and MiniMax-M3 match Blackwell, four of which correspond to GB200 and GB300, and he reproduced the initial weights on real B300 and rented GB200. He estimated that Kimi K3 and DeepSeek-V4-Pro each use 2,000 to 3,000 Blackwell GPUs, with a single-cluster procurement cost of about $150 million to $200 million, and said the results were independently verified by senior researchers at OpenAI and Anthropic.

  3. Unite.AISelected

    Unemployment Labs report: Weights of six Chinese open-source models show Blackwell initialization signatures

    Will DePue of Unemployment Labs published an analysis on November 12, 2026, stating that the initialization patterns of six open-weight models from DeepSeek, Moonshot AI, MiniMax, and Ant Group are consistent with NVIDIA's export-controlled Blackwell GPUs. Among them, DeepSeek-V4-Flash and MiniMax-M3 correspond to B200 or B300, while DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, and Ling-3.1-flash correspond to GB200 or GB300; the latest readable models from Alibaba, Zhipu, Tencent, Xiaomi, and StepFun instead show Hopper signatures. The report emphasizes that the match involves only the initialization stage, cannot prove which chips were used for subsequent training, and cannot determine chip ownership, location, or acquisition method, and therefore does not constitute a conclusion of smuggling or export-control violations.

Nov 12

  1. Hacker News TrendingSelected

    Unemployment Labs report claims that the weights of six Chinese models, including DeepSeek and Moonshot, carry...

    Will DePue of Unemployment Labs published a forensic report claiming that the public weights of six models—DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3, and Ant's Ling-3.1-flash—retain an initialization layout consistent with NVIDIA Blackwell chips (148 or 152 SMs), while earlier models in the same series correspond to Hopper. The method is based on the fact that when PyTorch uses Philox to generate initial weights, the thread count varies with the number of SMs. The author reproduced the retained weights of DeepSeek-V3, V4-Flash, and Kimi K3 on physical H100 and B300 machines and a rented GB200, and claims that all 24 models with publicly disclosed hardware match. The report emphasizes that this evidence only indicates the initialization hardware and cannot determine the chip's location, how it was obtained, or whether export controls were violated; AMD MI300X under a special configuration is the only alternative explanation that has not been fully ruled out.

Heat trend for this story

Current
6
Period peak
52 Nov 14 04:00
24-hour change
-80%

Move the pointer or select the chart to inspect hourly heat. Use the left and right arrow keys with a keyboard.

Related stories

  1. Related storiesDeepSeek open-sources inference framework DeepInfer
  2. Related storiesTestingCatalog claims that Jensen Huang has resigned from the position of NVIDIA CEO
  3. Related storiesNvidia announces Jensen Huang's resignation
  4. Related storiesAnthropic pitches investors and discloses revenue growth ahead of listing