All posts

Running open models on your own infrastructure

A practical look at deploying Gemma, Qwen, and Nemotron on-prem — for data privacy, predictable cost, and a clean cloud fallback.

Not every AI workload belongs in the cloud. If you're in finance, healthcare, or the public sector — or you just don't love the idea of your source code and customer data leaving the building — running open-weight models on infrastructure you control is often the better call. Here's how we think about it.

The models are good enough now

The open-weight field moved fast. Models like Gemma, Qwen, and Nemotron run comfortably on a single workstation or a modest GPU box and are more than capable for grounded, task-specific work: classification, extraction, retrieval-augmented answering, tool-calling inside an agent. You don't need a frontier model to wrap a mainframe API or triage a support queue — you need the right model, quantized and tuned for the job.

What you actually get

  • Data stays in-house. Sensitive code, documents, and PII never leave your network. With an air-gapped deployment, they never touch the public internet.
  • Predictable cost. Fixed infrastructure instead of metered per-token billing. Your AI bill doesn't balloon with usage or scale.
  • Control. You choose the model, the version, the update cadence. No surprise deprecations, no changing terms.

The pattern that works: local-first, cloud fallback

The mistake is treating it as all-or-nothing. The setup we deploy routes each request to a local model when the box is available, and falls back to a frontier cloud model automatically for the rare task that genuinely needs it. Most traffic is free and private; the hard cases still get answered.

Concretely, that's an inference runtime (Ollama or vLLM) serving the local model, a thin router that health-checks it and picks a backend, and the same MCP connectors underneath so agents don't care which model answered.

The unglamorous part: evals

The reason teams get burned by local models isn't capability — it's that they never measured. Before anything ships, we stand up evals against your actual tasks (not a leaderboard), benchmark a few candidate models, and pick on evidence. Then observability and rollback so it stays honest in production.

Done right, on-prem AI isn't a downgrade. It's more private, more predictable, and — for the workloads that fit — genuinely cheaper.

Curious whether your workload fits? Email us.

Have a legacy system you'd like to make agent-ready?

Book a call