Lawyer’s eight-V100 AI server moves DeepSeek from 2-bit to near-NVFP4 precision
A lawyer’s eight-GPU Nvidia V100 server now runs DeepSeek V4.1 Flash’s most-used experts at NVFP4 precision instead of 2-bit.
A lawyer in Columbia, South Carolina, named Nick Mermiges, has upgraded his private AI server to run DeepSeek V4.1 Flash at near-NVFP4 precision. The server, originally built on eight Nvidia V100 GPUs, now features enhanced RAM (1TB) and an RTX 3090. Mermiges, founder of Proprietary Legal Intelligence (PLI) and PLI Labs, used Claude Code to assist in the project's software development.
The upgrade was necessary due to crashes experienced in the earlier build, which were likely caused by a faulty power supply unit (PSU). The improvements include a 524K context, 99.1% next token prediction accuracy, and 0.079 nats KL divergence from Nvidia's reference model. However, the higher precision comes at a cost, with code output and legal writing now running 3% to 30% and 9% to 23% slower, respectively.
Written by urgent.news from Tom's Hardware's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
