Agency-wide document Q&A on AMD Instinct™ MI350P.Ask questions of your own records and get cited answers, on one LIQID UltraStack server, fully inside your enclave.
Keep document Q&A with one analyst team, or scale it to thousands inside the boundary.
Federal generative AI use cases went from 32 to 282 in a year.
A Q&A pilot that works for one team becomes agency-wide demand.
Eight of twelve agencies cite insufficient technical infrastructure.
Each server, switch and cable in the boundary must be authorized.
Classified and air-gapped networks can’t reach them.
Every answer has to link to the passage behind it.
One 10-GPU LIQID UltraStack server runs agency-wide document Q&A. A RAG system indexes your records and gpt-oss-120b answers with citations, on an open stack that installs with no internet.
gpt-oss-120b replicas, one full model per GPU, serve concurrent users.
Embedding and reranking models find the right passages.
Llama 3.1 8B parses and indexes new records as they arrive.
No cluster inside the boundary: one host, ten GPUs, no GPU network, and room to grow to 30.
Add GPUs and a new job, not a new cluster.
Scale document Q&A from one analyst cell to a full analyst team.
6 × gpt-oss-120b answers · 2 × embedding and reranking · 2 × Llama 3.1 8B ingestion
CIO / CDAO of agency or command. Pilot success; user demand beyond one team.
Classify, redact and draft release decisions with human sign-off.
10 × Llama 3.1 8B / Mistral Small 3.2 classify and PII · 6 × Llama 3.3 70B redaction rationale and drafting · 4 × gpt-oss-120b reviewer and QA
Records Officer / FOIA office. Statutory backlogs; litigation over delays.
Analyst Q&A, records release, translation and drafting in one accredited boundary.
10 × analyst Q&A · 12 × records release · 4 × Mistral Small 3.2 translation · 4 × gpt-oss-120b drafting assistants
Program office / mission owner. Sovereign AI mandates; many teams waiting for in-enclave AI.
Two questions settle the size of the opportunity: how many analysts want the pilot today, and how large is the record set behind it?
*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. Embedding and reranker models are selected per deployment from US/EU-origin open-weight options. This is a reference architecture, not an accredited system; authorization is the customer's responsibility. Federal AI data: GAO-25-107653.