Defensible e-discovery on AMD Instinct™ MI350P.Relevance, privilege and PII coding across millions of documents, on one LIQID UltraStack server, on a model version you control.
Send first-pass review to a vendor, or bring it in-house on a model you can explain.
First-pass budgets go to outside vendors, priced per document.
61% use AI-assisted review; under a third document protocols.
Model, configuration, validation and error rates.
A model that changes without notice is hard to defend.
Client-confidential data has to stay in your environment.
Every AI decision has to link back to its source document.
One 20-GPU LIQID UltraStack server runs the whole first-pass pipeline. Small models code every document, a 70B model drafts the privilege log and checks the calls, and a reviewer model reports error rates.
Mistral Small 3.2 and Llama 3.1 8B code every document for relevance, privilege and PII.
Llama 3.3 70B, one full model per GPU, drafts log entries and checks coding calls.
gpt-oss-120b reviews statistical samples and reports error rates.
No vendor, no cluster: one server, a pinned model and a validation record the court can follow.
Add GPUs and a new job, not a new cluster.
Every attorney in a practice group, plus dedicated capacity for deal diligence.
6 × Llama 3.3 70B interactive review · 4 × Llama 3.3 70B bulk diligence
Practice Group Leader / Director of Legal Innovation. Adoption beyond the pilot; deal volume.
Relevance, privilege and PII coding across millions of documents, on a version-pinned model.
12 × Mistral Small 3.2 coding fleet · 6 × Llama 3.3 70B privilege log and QC · 2 × gpt-oss-120b validation sampling
Litigation Support / e-Discovery Director. Large matter, review vendor cost, court scrutiny of AI.
Contract review, e-discovery, research Q&A and billing narratives on one host.
8 × Llama 3.3 70B contract review · 14 × e-discovery · 6 × gpt-oss-120b cited research Q&A · 2 × Llama 3.1 8B billing narratives
CIO / Chief Knowledge Officer. Client demand for AI with no data egress.
Two questions settle the size of the opportunity: how many documents are in your next large matter, and what will first-pass review cost?
*Efficiency targets compare against multi-server deployments of the same GPUs; they are LIQID internal projections using AMD published specifications, and results vary by configuration and workload.
Performance figures for the AMD Instinct MI350P are based on AMD published specifications and LIQID Phase-3 testing from 1 to 20 GPUs, and are subject to change. 30-GPU figures are projected from measured 20-GPU scaling. Every layout runs one model per GPU; 20- and 30-GPU layouts are sizing arithmetic on published model sizes, not benchmark results. Models shown are US and EU open-weight models. Outputs support, and do not replace, attorney review; review protocols and validation remain the firm's responsibility. Adoption data: EDRM 2025.