AI;DRWhat's happening in AI today?
AI;DRWhat's happening in AI today?

● ChipsNov 13, 2026 · 09:068 sources

Forensic Report Finds No Nvidia Blackwell Signature in Newest DeepSeek V4.1 Flash Model

From marketbrief.now

AI;DR

A forensic report found no Nvidia Blackwell hardware signature in DeepSeek-V4.1-Flash. Both V4.1-Flash and LongCat were trained on Huawei Ascend chips. DeepSeek's open-source FlashMLA repository listed support for Nvidia SM100 and SM103 Blackwell architectures.

Read originalSign in to suggest
Image: marketbrief.now

Sources

+ 7 more
  1. SUPPORT@willdepue“This is that search on DeepSeek-V4-Flash. Every processor count gives noise until 148 (Blackwell), where the weights match.”x.com
  2. SUPPORT@nrehiew_“Even if weights are changed during training, their relative ordering from initialization should be preserved and you can reverse-engineer the SM count of the GPU that produced the initial weights”x.com
  3. SUPPORT@dummerspast39“His own report, buried in the appendix: "We recovered no signature from DeepSeek-V4.1-Flash (after an exhaustive search)"”x.com
  4. SUPPORT@teortaxestex“we've already estimated that DeepSeek's inference server for V4.1-Flash is about 2.7K-3K GB200s. V4-Pro was a flop… It seems they didn't grow much. 3K Blackwells is like 6 MW. This is utterly insufficient.”x.com
  5. SUPPORT@nicbstme“Like in their repo FlashMLA README lists “NVIDIA SM100 / SM103 GPU” (Blackwell) and they even deleted the support for Hopper at the end of Sept.”x.com
  6. SUPPORT@deanwball“When I first saw this I thought, “wow, this is dna evidence,” and then it quickly dawned me that this is in fact orders of magnitude stronger than literal dna evidence submitted in court.”x.com
  7. SUPPORT@hrkrshnn“Deepseek has open-source kernels for Blackwell. This shouldn't be a surprise; they and everyone else in the world have access to pretty much any chips.”x.com
#deepseek#llmScore 0.33.2k tokensMarkdown
Read original