Technology
Lawyer upgrades private AI server, upgrading DeepSeek to near-NVFP4 precision
A trial lawyer has upgraded his custom private AI server built with used Nvidia V100 GPUs, moving DeepSeek V4.1 Flash to near-NVFP4 precision while adding 1TB of RAM and an RTX 3090.
A South Carolina trial lawyer running a private AI server with eight used Nvidia V100 GPUs has updated his setup. The system now runs DeepSeek V4.1 Flash at near-NVFP4 precision for about 80% of its workload, moving up from 2-bit precision. Alongside this change, the system received upgrades to 1TB of RAM and an RTX 3090. While the precision upgrade improved stability and eliminated previous crashes, it came at a performance cost. Code output speeds dropped from a range of 160–215 tokens per second down to 150–180 tokens per second.
Powered by AI