Lawyer runs DeepSeek Flash at near-NVFP4 quality on eight 2018 V100s: 524K context, 100+ tok/s, fully local

TumbleweedNew6515 · reddit · 2026-11-13

A lawyer known for hoarding GPUs published his fifth local-deployment build: DeepSeek Flash 1017 running on eight V100-SXM2-32GB cards at near-NVFP4 quality, with 524K context, 118.9 tok/s single-request decode, and 159 tok/s with two concurrent chats — fully local, powering his legal workflow.

The Volta Max uPLIft approach

He also runs two DGX Sparks handling 3-4 concurrent DeepSeek agent sessions at 1M context, plus 8 more V100s and a 512GB ECC RAM server for expansion. Notably, newer models (Opus 5.5, Fable 5.2) now handle the configuration, custom Volta kernels, and activation-guided expert requantization automatically — the first time his frontier-built legal harness has been fully replaced by a local stack. Code and benchmarks are open-sourced on GitHub.

Original post →

More from coding & agent

coding & agent channel →