Nick Mermiges, founder of PLI Labs, has upgraded his private AI server, which utilizes eight Nvidia Tesla V100 GPUs. The system now runs DeepSeek V4.1 Flash with near-NVFP4 precision by dynamically swapping experts between 2-bit and 4-bit formats.
The hardware configuration was updated with 1TB of RAM, replacing the previous 256GB setup. Mermiges also added an Nvidia RTX 3090 to the system and resolved stability issues by replacing a faulty power supply unit.
Performance metrics show that 78% to 86% of model decode routes now utilize NVFP4 experts. While this increases fidelity, it results in a 3% to 30% decrease in generation speed for code and legal text compared to the previous 2-bit build.
The updated server supports multiple simultaneous requests, achieving a total output of 159 tokens per second when running two concurrent tasks. Mermiges plans to expand the infrastructure by adding a second server with eight additional V100 GPUs.
