GoKawiilLawyer upgrades homemade V100 AI server, moves DeepSeek model to 4-bit precision
Nick Mermiges, a South Carolina trial lawyer and founder of PLI Labs, upgraded his private eight-GPU V100 AI server by adding 1TB of RAM and an RTX 3090, shifting roughly 80% of its workload from 2-bit to 4-bit NVFP4 precision. The server runs DeepSeek V4.1 Flash with a custom legal workflow he built using Claude Code, and the upgrade traded some speed—code output dropped from 160–215 to 150–180 tokens per second—for stability, after crashes traced to a faulty power supply were resolved.
GoKawiil's interpretation of the reporting above, not reported fact.
The project illustrates how individual professionals, not just large labs, are building on-premises AI systems to keep sensitive data like legal documents off external servers, which Mermiges says never leave his office. It also suggests aging, discounted datacenter GPUs like the V100 can still be pressed into serious AI work with enough custom engineering, though reliance on a soon-to-be-unsupported driver branch points to the setup's limited long-term viability. His pivot from a technology consultancy to offering private LLM hosting as a service hints that this kind of DIY on-prem AI could become a niche business model for privacy-conscious clients.
- A lawyer upgraded his 8x Nvidia V100 server with 1TB RAM and an RTX 3090 to run DeepSeek V4.1 Flash at near-4-bit precision.
- The upgrade improved stability by fixing a power-supply issue but slightly reduced code-generation speed.
- The builder, Nick Mermiges, now offers private on-premises LLM hosting through his consultancy PLI Labs.
NVIDIA RTX 3090 Graphics Card — This is literally the card the lawyer added to his homebrew AI server to boost precision and performance. With 24GB of VRAM, it's a favorite among DIY AI builders for running large local models like DeepSeek alongside older datacenter GPUs.ゼ
See NVIDIA RTX 3090 Graphics Card on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: tomshardware.com — Shane Downing, 2026-11-14
Published there as: “Lawyer’s eight-V100 AI server moves DeepSeek from 2-bit to near-NVFP4 precision”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.