Six Chinese Models Show Blackwell Fingerprints

// 1h agoRESEARCH PAPER

Six Chinese Models Show Blackwell Fingerprints

An Unemployment Labs report claims six open-weight models from DeepSeek, Moonshot AI, MiniMax, and Ant Group preserve initialization patterns matching NVIDIA Blackwell hardware. Physical reproductions support the fingerprinting method, but the evidence indicates hardware access—not where training occurred or whether chips were illegally transferred.

// ANALYSIS

If independently validated, this turns model checkpoints into hardware-provenance evidence and complicates export controls.

–Identified models include DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3, and Ling-3.1-Flash.
–The analysis searched more than six trillion GPU-layout and seed combinations, with matches reportedly far beyond random-chance expectations.
–Remote access through overseas data centers remains a plausible explanation; it does not require Blackwell systems to be physically shipped into China.
–Developers and policymakers may gain a new way to audit training provenance, but the technique needs independent replication and peer review before it can support enforcement decisions.

// TAGS

checkpoint-forensicsllmopen-weightsgputraining-infraresearch

DISCOVERED

1h ago

2026-11-12

PUBLISHED

1h ago

2026-11-12

RELEVANCE

9/ 10

AUTHOR

theo

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

UPDATE37m ago

Framer Agent Generates Images, Shaders

Framer’s AI Agent can now turn natural-language prompts into generated imagery and interactive shader effects, including pointer-driven smudges, halftones, CRT visuals, bokeh, and pencil styles. The workflow aims to produce polished web interactions without manual code or After Effects.

OPEN SOURCE44m ago

Perplexity Decider v1.3 Adds Locate Extraction

Perplexity open-sourced its latest decision models under Apache licensing, adding Locate for extracting exact text spans and character positions. Both models are free on OpenRouter for one week and support classification, routing, moderation, reranking, and structured evaluation.

LAUNCH49m ago

Daybreak Red Team automates continuous pen tests

OpenAI’s Daybreak Red Team puts coordinated AI agents to work testing web applications, validating vulnerabilities, and checking whether patches actually stop exploits. The early-access service targets organizations that need continuous offensive security coverage without maintaining a dedicated red team.