DeepSeek, Moonshot, MiniMax, Ant Group AI models show NVIDIA Blackwell chip signatures
Forensic analysis of published model weights indicates DeepSeek, Moonshot AI, MiniMax and Ant Group used export-controlled NVIDIA hardware.
Key points
- Six models released since April 2026 contain numerical patterns matching NVIDIA Blackwell hardware.
- Researchers confirmed the hardware match on physical chips for DeepSeek-V4-Flash and Kimi K3.
- The findings do not prove export controls were violated or reveal where the chips were located.
Forensic analysis of publicly released AI models indicates that DeepSeek, Moonshot AI, MiniMax and Ant Group have accessed NVIDIA’s Blackwell GPUs, according to a report by Unemployment Labs.reported The study identifies numerical patterns in model weights that align with Blackwell hardware configurations, suggesting these companies are using the chips to create their newest AI models despite export restrictions.The study finds that six Chinese AI models contain digital traces of restricted NVIDIA hardware.Weight forensics reveal Blackwell SM counts in six models, with physical validation for two.
Evidence: reported
Chinese frontier AI models carry the fingerprint of export-controlled GPUsaifleet-hn-ai
How we got here
Prior to this technical evidence, U.S. officials and press reports had alleged that DeepSeek and Moonshot AI were using Blackwell chips without publishing supporting data.reported Earlier models from these same companies matched NVIDIA’s older Hopper generation hardware. The new analysis provides the first technical evidence of access to this advanced computing hardware through the models’ own published weights.
What is new
The report identifies six specific models released since April 2026 that carry Blackwell fingerprints.reported DeepSeek-V4-Flash and MiniMax-M3 match B200 or B300 chips, which feature 148 processor cores. DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash match GB200 or GB300 chips, which have 152 processor cores. Researchers confirmed the match on physical Blackwell chips for DeepSeek-V4-Flash and Kimi K3.
Evidence: reported
Chinese frontier AI models carry the fingerprint of export-controlled GPUsaifleet-hn-ai
What to watch
The analysis excludes specially configured AMD hardware for DeepSeek-V4-Flash and considers it very unlikely for the other five models.reported However, the findings do not establish where the chips were located or how access was obtained. The report states that the fingerprint supports, without proving, the inference that Blackwell also performed the training.
Evidence: reported
Chinese frontier AI models carry the fingerprint of export-controlled GPUsaifleet-hn-ai
Availability and limits
Developers using these models can observe hardware-specific numerical patterns in the published weights. The study notes that no U.S. export licenses for these chips to China have been reported.reported The evidence comes entirely from the models’ own weights, not from leaks. Kimi K3.1-Pro may be a checkpoint of K3 with further post-training, starting from the same seed.
Why it matters
- Developers can verify hardware origins by examining numerical patterns in model weights, offering a new forensic method.
- The presence of Blackwell signatures suggests Chinese labs bypassed or obtained restricted hardware, impacting competitive assessments.
- Physical confirmation on DeepSeek-V4-Flash and Kimi K3 adds credibility to the weight-based analysis method.
- The inability to distinguish H800 from H100 or H200 in Hopper models limits retrospective analysis of older releases.
Topics: models, hardware, forensics, NVIDIA, export controls
Updated
Then / now / next
Earlier models from these companies matched NVIDIA’s older Hopper generation hardware.
Six models released since April 2026 contain numerical patterns matching Blackwell hardware.
The findings do not establish where the chips were located or how access was obtained.
Story timeline
- Then Earlier models from these companies matched NVIDIA’s older Hopper generation hardware.
- 12 Nov 2026 Chinese frontier AI models carry the fingerprint of export-controlled GPUs
- Now Six models released since April 2026 contain numerical patterns matching Blackwell hardware.
- Next The findings do not establish where the chips were located or how access was obtained.
For readers
Developers
- Can examine published weights for hardware-specific numerical patterns.
- Should note that AMD hardware was excluded for one model.
Researchers
- Can use physical chip confirmation as a benchmark for weight analysis.
- Should consider that training inference is supported but not proven.
Policy
- No U.S. export licenses for these chips to China have been reported.
Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
Which models show Blackwell GPU signatures?
DeepSeek-V4-Flash, MiniMax-M3, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash show matching patterns.
Does this prove export control violations?
No, the findings provide technical evidence of access but do not establish whether export controls were violated.
How was the hardware match confirmed?
Researchers recovered traces of initial values and confirmed the match on physical Blackwell chips for two models.
Connected stories
Reading receipt
- Six models released since April 2026 contain numerical patterns matching NVIDIA Blackwell hardware.
- Researchers confirmed the hardware match on physical chips for DeepSeek-V4-Flash and Kimi K3.
- The findings do not prove export controls were violated or reveal where the chips were located.
Still open: Where were the chips located?
Still open: How was access to the restricted hardware obtained?