DeepSeek’s latest firmware upgrade has pushed a lawyer’s eight‑GPU AI server from a 2‑bit to a near‑NVFP4 precision, dramatically improving performance while keeping documents in‑office.
Hardware Overhaul
- Eight Nvidia V100 GPUs upgraded to 4‑bit NVFP4 precision.
- 1TB of RAM added, up from 256GB.
- RTX 3090 integrated for additional compute headroom.
Performance & Stability Gains
| Metric | Before | After |
|---|---|---|
| Tokens per second | 160–215 | 150–180 |
| Crash incidents | Frequent (PCIe suspected) | None (PSU issue resolved) |
| Precision level | 2‑bit | Near‑NVFP4 (4‑bit) |
Software & Model Integration
The upgrade leverages DeepSeek V4.1 Flash, built from expert kernels that fit the 4‑bit format. Claude Code authored custom kernels, while Opus 5.5 and Fable 5.2 handled re‑quantization. The system now runs a legal harness that keeps all documents on‑prem, meeting the lawyer’s multi‑request requirement.