Private AI — Hermes
A capable assistant model that never phones home.
Hermes is a family of open-weight assistant models built for instruction following and tool use — and unlike the big hosted assistants, it runs entirely on hardware you control. We size it, serve it, wire it into your tools, and keep it tuned.
Why Hermes
Open weights, closed network.
The Hermes model family is published with open weights, which means the actual model — not a rented API — can run on your GPU node. Prompts, documents, and responses never cross your network boundary. Open-weight assistants trail frontier online models slightly on raw capability; for work that touches customer records, contracts, and HR data, that trade is exactly the one you want to make.
Model selection and sizing
Which Hermes variant fits your workload and your GPU — we match parameter count, quantization, and context length to the jobs you actually have, not the biggest number on a leaderboard.
Serving stack
Served through Ollama or a high-throughput runtime like vLLM, hardened, with the local API reachable only from your agent VLAN — not the office coffee machine.
Your documents, in context
We wire retrieval over your own knowledge base — policies, wikis, tickets — so the assistant answers from your material, with citations, all of it local.
Agent-ready
Hermes models are tuned for tool use, which makes them a natural brain for OpenClaw agents — one GPU node can power both.
Pilot before you commit
We run a two-week pilot on real (or realistic) tasks and show you quality against the hosted assistant your team already uses. Numbers first, hardware second.
Tuned and current
New Hermes releases land regularly. We evaluate them against your pilot benchmarks and upgrade only when they're genuinely better — with a rollback path.
How we set it up
From workload audit to working assistant.
Audit the workload
What should the assistant do — draft, summarize, answer from documents, triage tickets? Task mix drives model size, context needs, and hardware.
Typical: 1 week
Size the hardware
VRAM math against real models: we specify the GPU node that carries your workload with headroom — or tell you the office server you already own will do.
Typical: 3–5 days
Serve the model
Ollama or vLLM installed and hardened, Hermes weights pulled and quantized, API exposed on the AI VLAN only, load-tested against expected concurrency.
Typical: 1 week
Wire it in
Retrieval over your documents, connections to OpenClaw or your tools, a simple interface your team will actually use. Pilot group starts work.
Typical: 1–2 weeks
Evaluate and operate
Quality scored against the pilot benchmarks, usage monitored, model updates staged — and every request logged as SOC 2 evidence.
Ongoing
Hosting
Your model, our rack.
A Hermes node in our data center gets you GPU hardware without the noise, heat, or UPS anxiety: A+B power, N+1 cooling, and the NOC watching inference health around the clock. Your staff reaches it over an encrypted tunnel; the model and its context never leave the rack. See colocation pricing or build a hosting quote — a single GPU node typically lands in a quarter or half rack.
Find out what your data can do at home.
Bring us three tasks your team does weekly. We'll show Hermes doing them on private hardware — and tell you honestly where it falls short of the cloud.