Skip to content
View ProprietaryLegal's full-sized avatar

Block or report ProprietaryLegal

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
ProprietaryLegal/README.md

PLI Labs — Proprietary Legal Intelligence

Your proprietary work product is your proprietary intelligence.

Nick Mermiges, Esq. is a trial lawyer in Columbia, South Carolina, with more than eleven years of civil litigation, trials and appeals, and the founder of Proprietary Legal Intelligence. He runs his own litigation practice on AI every day and helps small and mid-sized civil litigation firms find where their work backs up and fix it. PLI Labs publishes the open research from that work: everything here was built first inside a working trial practice, on hardware he owns.

Website · Claude workflows · Local models · PLI Labs · Research · Hugging Face · Contact

In the press: Tom's Hardware, November 9, 2026, by Shane Downing, on the eight-V100 build: Lawyer builds private AI server with eight used Nvidia V100 GPUs — up to 215 tokens per second to keep legal files offline, costing about $2,400 per four-GPU board.

The builds

Every build serves DeepSeek-V4.1-Flash-1017 on hardware you can own. Each card gives that build's own MEASURED tokens per second for decode, prefill and concurrent requests (at 128K tokens of context where the build publishes that row) and its KV cache; its repository gives the method, the conditions and the losses beside the wins.

Every routed expert at NVIDIA's NVFP4 precision on twelve cards, three pipeline stages of four.

Volta 12X AmPLIfy: 12x V100-SXM2-32GB; at 128K context: decode 112 tok/s for one request after a turn, 124 after a cold read; prefill 4,426 tok/s, 4,405 at 248K, 4,223 at 426K; four requests together 312 tok/s in sum; KV cache: 524K tokens of context per request, four slots

As much of the routed work at NVIDIA's NVFP4 precision as fits on eight cards, every expert on the cards.

Volta Max uPLIft: 8x V100-SXM2-32GB; at 128K context: decode 148 tok/s on code, 146 on JSON, 79 on legal text, 77 on prose; prefill 2,727 to 3,145 tok/s, first token after 43.8 s; three requests together 176 tok/s, same tokens as each alone; KV cache: 524K tokens of context per request, two slots, a follow-up resumes in 0.69 s

The whole 249 GB model resident on eight 2018 cards.

Volta sPLIt: 8x V100-SXM2-32GB; at 128K context: decode 183 tok/s on code, 197 on JSON, 111 on legal text, 91 on prose; prefill 2,200 tok/s, first token after 59 s; one request at a time; KV cache: 524K tokens of context per request

Every routed expert at exactly NVFP4 precision on eight cards, the less-used ones served from host RAM.

Volta Full NVFP4: 8x V100-SXM2-32GB plus host RAM; decode 28 to 30 tok/s on legal-style text; prefill 1,419 to 1,595 tok/s, a 498K-token prompt read in 326.6 s; one request at a time; KV pool 2.08M tokens, 524K context per request

Two DGX Sparks serve the model; an x86 machine holds the Engram tables in RAM over RDMA and runs the drafter on two RTX 3090s.

DGX sPLIce: 2x DGX Spark, 2x RTX 3090 and one x86 host; at 128K context: decode 62 tok/s on prose for one stream, prefill 2,650 tok/s; four streams 322 tok/s on JSON, 192 on code; KV pool 1.93M tokens, 327,680 tokens of context per request

The research

The long read and the source research behind the Volta builds: the card's architecture, every optimization build by build (with the ones that lost), the test method, kernel building for sm_70, and a hardware list with links for building your own boards. Repository.

The skill

An agent skill for Claude Code and Codex that runs the campaign behind these builds on your own cards: intake, upstream first, layout arithmetic, a test harness with strict judges, a measured time budget, a gated hill-climbing loop that keeps only what passes, kernels, independent review, publishing. Install with ./install.sh --both.

More

Components, earlier builds and mirrors of upstream work are under all repositories.

PLI Labs is the open research arm of Proprietary Legal Intelligence. We publish methods, code and measured results so you can inspect the work before hiring us. For AI systems on approved frontier APIs, private cloud or hardware you own, reach us through the contact page.

Proprietary Legal Intelligence LLC is a technology consultancy, not a law firm, and does not provide legal services or legal advice.

Popular repositories Loading

  1. Volta-sPLIt-DeepSeek-V4.1-Flash-1017-8xV100-SXM2-PP2xTP4 Volta-sPLIt-DeepSeek-V4.1-Flash-1017-8xV100-SXM2-PP2xTP4 Public

    DeepSeek-V4.1-Flash-1017 on 8x V100-SXM2 (2018): 119.0 tok/s decode, 120.8 thinking on (measured), 524K context, 300K-token first token in 178 s. PP2xTP4 port of 1Cat-vLLM #1277 with every patch, m…

    Python 36 4

  2. Volta-Max-uPLIft-FP4-DeepSeek-V4.1-Flash-1017-NVFP4-8xV100-SXM2-PP2xTP4 Volta-Max-uPLIft-FP4-DeepSeek-V4.1-Flash-1017-NVFP4-8xV100-SXM2-PP2xTP4 Public

    Volta Max uPLIft: DeepSeek-V4.1-Flash-1017 on 8x V100-SXM2 (2018) with as much of the routed work at NVIDIA's NVFP4 precision as fits and every routed expert on the cards. One Hugging Face package,…

    Python 21

  3. DGX-sPLIce-DeepSeek-V4.1-Flash-2xDGX-Spark-2xRTX3090-drafter-x86-offload DGX-sPLIce-DeepSeek-V4.1-Flash-2xDGX-Spark-2xRTX3090-drafter-x86-offload Public

    DGX sPLIce: DeepSeek-V4.1-Flash on 2x DGX Spark + 2x RTX 3090 drafter, x86 offload (Engram over RDMA, DSpark drafter). Prefill 2,606-2,827 tok/s to 256K; 322 tok/s structured, 192 tok/s code at 4 s…

    Python 6

  4. deepseek-v4-flash-v100 deepseek-v4-flash-v100 Public

    DeepSeek-V4-Flash and V4.1-Flash on NVIDIA V100: llama.cpp recipe and quantized-KV fixes (V4); 1Cat-vLLM PP2xTP4 results so far and patches (V4.1). PLI Labs.

    Python 5 2

  5. ProprietaryLegal ProprietaryLegal Public

    PLI Labs — Proprietary Legal Intelligence: open legal-AI research, legal model prunes, owned-hardware serving recipes and tools. proprietarylegal.ai · PLI Labs

    2

  6. pli-labs pli-labs Public

    PLI Labs: legal model prunes (REAPs), owned-hardware serving research, evaluation methods, and open-source tools from Proprietary Legal Intelligence.

    Python 2