“Local” and “offline” are different properties

A local language model may run inference on a company server while the rest of the software stack remains connected. A download client can contact a repository for files or metadata; an application may use cloud search, analytics, or an external integration; an update or monitoring service may make its own requests.

Before a pilot, map the data and network flows: model downloads, inference, the knowledge base, user interface, telemetry, updates, backups, and tools. For each route, record an owner, destination, data type, and reason for the connection.

Settings reduce requests but do not replace controls

Hugging Face documentation describes `HF_HUB_OFFLINE`, which disables Hub client network calls and uses local files; without the required files in cache, a download cannot complete. Ollama and other runtimes also document separate options for cloud features or usage statistics. Names and behavior depend on the product version, so check the exact build you plan to deploy.

Turning off telemetry in one component does not block other processes. An environment variable does not create a network perimeter and cannot prevent an application from contacting a new address after an update. If the requirement is “data must not leave the environment,” enforce it at the network layer, not only in a configuration setting.

A three-stage check

1. **Inventory.** List services and external addresses needed for installation, downloads, inference, updates, and diagnostics. Separate required, conditionally permitted, and prohibited traffic.
2. **Isolation.** Allow outbound connections through an allowlist, and centrally control DNS and direct internet access. Check containers, background tasks, and operator workstations separately.
3. **Observation.** Capture network events during a cold start, routine request, error, restart, and update. Check not only hostnames but actual IPs, DNS, proxies, and log contents.

Repeat the test after changing an image, model, or application version. If the environment must be fully offline, prepare and test a portable cache in advance instead of disconnecting the network on launch day.

Data and access

“Where does a request go?” and “who can read it inside?” are separate questions. A local server does not remove the need for authentication, access controls, logging, and retention rules. Hidden system logs and error dumps can also contain prompt fragments.

Use anonymized documents and test accounts for the pilot. Define which information the model may access, what must not be indexed, and who approves exceptions. Write a testable criterion, such as no outbound connections during inference other than explicitly listed monitoring.

Practical next step

Ask the deployment team for a network-flow diagram and a test protocol covering startup, inference, errors, and updates. The network administrator can then verify the “offline” claim against observable logs. This is a quick, low-cost check before connecting sensitive business documents.