Skip to content
Frontier Leaf FL markFLFrontierLeaf

A field journal of new AI developments.

Model updates

Model weights hint at Blackwell hardware use by Chinese AI labs

Forensic analysis of six recent models from firms including DeepSeek and Moonshot AI shows numerical patterns matching NVIDIA’s export-controlled Blackwell chips.

Depth
1 outlet · newest source 85h ago
Model weights hint at Blackwell hardware use by Chinese AI labs

Key points

  • Six models released since April 2026 carry numerical patterns matching NVIDIA Blackwell hardware.
  • Earlier models from the same developers matched the older Hopper generation.
  • The findings support, but do not prove, that Blackwell chips were used for training.

Publicly released AI models from four Chinese companies contain numerical fingerprints consistent with NVIDIA’s Blackwell GPUs. This forensic finding suggests these developers have accessed hardware that faces strict export controls to the region.This analysis suggests Chinese AI firms are using restricted NVIDIA chips to create new models, based on patterns in the code.Weight initialisation traces in six recent models align with Blackwell SM counts, offering technical corroboration for prior allegations of restricted hardware access.

The evidence

A report from unemploymentlabs.com shows that DeepSeek, Moonshot AI, MiniMax and Ant Group released six models with patterns matching Blackwell chips. These patterns appear in models released since April 2026, whereas earlier versions from the same teams matched NVIDIA’s older Hopper generation.

Evidence: reportedChinese frontier AI models carry the fingerprint of export-controlled GPUs

aifleet-hn-ai

The researchers recovered traces of initial values from the published model weights. They then searched for hardware configurations that could reproduce these specific numerical signatures. For DeepSeek-V4-Flash and Kimi K3, they confirmed the match on physical Blackwell chips.

DeepSeek-V4-Flash and MiniMax-M3 match the B200 or B300 chips, which have 148 processor cores. DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash match the GB200 or GB300 chips, which have 152 cores. Kimi K3.1-Pro may be a checkpoint of K3 with further post-training.

The decision

The evidence comes entirely from the models’ own published weights, not from leaks or insider accounts. The fingerprint identifies how the weights were created and supports the inference that Blackwell performed the training, without proving it.

Specially configured AMD hardware is the only alternative identified, as it is also restricted from export. The researchers excluded AMD for DeepSeek-V4-Flash and consider it very unlikely for the other five models.

Evidence: reportedChinese frontier AI models carry the fingerprint of export-controlled GPUs

aifleet-hn-ai

The takeaway

These findings provide technical evidence of access to advanced computing hardware. However, they do not establish where the chips were located, how access was obtained, or whether export controls were violated.

U.S. officials and press reports had previously alleged that DeepSeek and Moonshot AI were using Blackwell chips, but they did not publish technical evidence.reported This analysis provides technical evidence of access to advanced computing hardware.

Building on this

Developers evaluating these Chinese models should note the hardware signatures embedded in their weights. The presence of Blackwell fingerprints may influence assumptions about model performance characteristics derived from that specific architecture.

Teams building applications on these weights must recognise that the hardware origin remains partially inferred. The analysis confirms the weight initialisation matches Blackwell, but does not prove the full training process occurred on that hardware.

Why it matters

  • Technical verification of hardware access allows IT buyers to assess the provenance of open-weight models from Chinese labs.
  • The distinction between Hopper and Blackwell signatures reflects the hardware generations used in successive model releases.
  • The inability to prove training location or export violations introduces uncertainty for compliance-focused procurement teams.
  • Confirmation on physical chips for two models provides direct evidence for the hardware match in those specific cases.

Topics: AI models, NVIDIA, export controls, hardware forensics, DeepSeek

Updated

Then / now / next

Then

Earlier models from these developers matched NVIDIA’s older Hopper generation hardware.

Now

Six models released since April 2026 carry numerical patterns matching Blackwell hardware.

Next

Observers will watch for further technical evidence clarifying how access to restricted chips was obtained.

Story timeline

  1. Then Earlier models from these developers matched NVIDIA’s older Hopper generation hardware.
  2. 12 Nov 2026 Chinese frontier AI models carry the fingerprint of export-controlled GPUs
  3. Now Six models released since April 2026 carry numerical patterns matching Blackwell hardware.
  4. Next Observers will watch for further technical evidence clarifying how access to restricted chips was obtained.

The strongest doubt

Counterpoint from the sources

The analysis supports the inference of Blackwell use but explicitly states it does not prove the training process occurred on that hardware or that regulations were broken.

For readers

Developers

  • Model weights contain hardware-specific numerical patterns that may reflect the training environment.
  • Two models were confirmed to match Blackwell chips on physical hardware tests.

Policy

  • Technical evidence exists for access to restricted chips, but location and method remain unknown.
  • No U.S. export licenses for these chips to China have been reported.

Researchers

  • Forensic methods can recover hardware traces from published model initial values.
  • AMD hardware remains a theoretical alternative but is unlikely for most models analysed.

Ask the article

Local keyword search of this article; no generated answers.

Did this article give you the evidence you needed?

No vote saved on this device.
Your choice stays on this device. No public tally.

Questions readers ask

Which specific models show Blackwell fingerprints?

DeepSeek-V4-Flash, DeepSeek-V4-Pro, MiniMax-M3, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash show patterns matching Blackwell chips.

Does this prove export control violations?

No, the analysis states it does not establish whether export controls were violated or how access was obtained.

Could AMD hardware have been used instead?

Specially configured AMD hardware is the only alternative, but it was excluded for DeepSeek-V4-Flash and deemed very unlikely for the others.

Connected stories

This storyNvidia’s Huang steps down as AI startup backers weigh lost allyNvidia’s Huang steps down as AI Nvidia GR00T 2 reduces robot teleoperation via video-based world-action modellingNvidia GR00T 2 reduces robot telNvidia CEO Huang's affair with VP Powell adds to spy scrutinyNvidia CEO Huang's affair with VGPT-6 Astra benchmarks suggest general models may replace specialized robotics AIGPT-6 Astra benchmarks suggest g

Reading receipt

  • Six models released since April 2026 carry numerical patterns matching NVIDIA Blackwell hardware.
  • Earlier models from the same developers matched the older Hopper generation.
  • The findings support, but do not prove, that Blackwell chips were used for training.

Still open: How did these companies obtain access to Blackwell chips if no export licenses were reported?

Still open: Did the actual training process occur on Blackwell hardware, or only parts of it?

Time spent: 0:00 · Depth: Full

Sources

  1. Hacker News: AI posts over 50 points

Related stories

Background reading