Home AI Server Runs DeepSeek Model on 12 GPUs

A private upgrade adds a third board of used Nvidia V100 cards so every expert of a DeepSeek model runs at 4-bit precision.

Updated on Nov. 20, 2026 in Artificial Intelligence

Close view of parallel bare green circuit boards with exposed silver heatsinks and copper cabling on a pale wooden workbench in soft daylight.
Nick Mermiges expanded his private AI server to 12 used Nvidia V100 GPUs for $9,450, enabling every expert of a DeepSeek model to run at 4-bit precision. AI Illustration. Upload story photo >

Nick Mermiges upgraded his private AI server from eight to 12 used Nvidia V100 GPUs for $9,450, allowing every expert of the DeepSeek V4.1 Flash model to run at 4-bit NVFP4 precision. The setup has been his default DeepSeek machine since Monday.

Why it matters

The upgrade shows that older datacenter GPUs can host a large mixture-of-experts model locally without compressing every expert further. Running the full model on consumer-accessible hardware reduces reliance on cloud services for long-context work.

The upgraded server reaches about 184 tokens per second on code decoding, with 12 V100 cards capped at 175 W and 1,452 MHz each. The V100 lacks FP4 or FP8 tensor cores, so the NVFP4 experts run without native hardware acceleration for that format.

The players

Nvidia

Nvidia is a American semiconductor company whose GPUs are widely used for artificial intelligence training and inference.

DeepSeek

DeepSeek is a Chinese artificial intelligence company that develops open-weight large language models.

The details

The 15,360 experts now sit at full precision across three boards, with layers split 0-13, 14-27 and 28-39. Cards on each board share NVLink, while transfers between boards cross PCIe through two switches on the host.

Timeline

  1. Since Monday: the setup has served as Mermiges' default DeepSeek machine.

The Tech Race

The setup relies on Nvidia's NVFP4 format, which Nvidia introduced for its newer Blackwell architecture, rather than the older V100 hardware it was run on. The workaround shows how hobbyist builds on previous-generation datacenter cards are testing the reach of new low-precision formats.

Readers who run local AI models can see that used datacenter GPUs may support large mixture-of-experts models at high precision. The approach requires significant upfront hardware spending and technical setup before it provides any advantage over cloud access.

The takeaway

Hobbyists with older datacenter cards can run full-precision expert sets on local hardware by splitting layers across boards. Anyone considering this route should weigh the cost of the GPUs and the setup time against cloud rental rates.

Further reading

Follow more home-lab and local-model coverage in the Artificial Intelligence section.

Source note: This article includes information reported by Tom's Hardware.