Is Your Local AI Really Offline? A Practical Test Guide

Local inference means a model can run on your machine. An offline workflow means the complete task can finish without reaching a remote service. A desktop application can satisfy the first description while still using cloud search, remote embeddings or online tools for parts of the work.

This guide provides a proposed test protocol, reviewed September 14, 2026. It is not a benchmark of an installed workstation. Use non-sensitive sample files and a test environment, then record exactly which task and configuration you evaluated.

Write the boundary before choosing a tool. Does the requirement cover prompts and answers only, or also documents, embeddings, logs and tool calls? Is communication with another machine on the office network allowed? A workflow can be disconnected from the internet while still accessing a local network service.

Ollama’s FAQ distinguishes local operation from cloud-hosted models and documents a local-only configuration. This is an example of why the product name alone is insufficient evidence: the same application can expose more than one execution mode.

A hardware purchase should follow the workload, not replace this boundary analysis. More RAM or a faster GPU cannot make an external search API operate offline. Our local workstation economics guide addresses cost after the dependency requirements are understood.

Trace a single representative task from input to saved result. The table separates setup dependencies from dependencies that may be used every time. Mark each row local, remote, unused or unknown only after checking the selected configuration.

A desktop window is an interface, not evidence of where computation occurs. Similarly, a file stored locally can be uploaded by a connector. Inspect the actual tool configuration and data path rather than inferring locality from where the user clicked.

The reviewed Ollama documentation provides disable_ollama_cloud in its server configuration and the alternative environment variable OLLAMA_NO_CLOUD=1. It says to restart Ollama after changing the configuration and identifies a log message confirming cloud features are disabled. Follow the current documentation for the installed version.

That setting disables Ollama cloud models and web search. It does not establish that a separate desktop frontend, browser extension or tool server has stopped making its own network requests. Treat each connected component as a separate dependency in the assessment.

Before applying a configuration change, save the existing setting and identify how to restore it. Perform the experiment on your own test environment without interrupting other people’s work. A narrowly scoped test is easier to interpret than changing a machine’s network policy while unrelated applications are active.

Prepare a harmless document with distinctive text and a clear expected result. Download the intended local model in advance. Then record the runtime version, selected model, enabled tools and which network boundary you will disconnect. Avoid using a cached answer as the test case.

Run the same task online and offline, saving outputs and errors. Use fresh sample content so successful completion requires processing the test input. If you inspect network activity, distinguish attempted connections from successful transfers; a failed request may still reveal a dependency.

The protocol below is a proposed assessment sequence. It is not evidence that we performed the experiment. A useful record should be repeatable by someone else with the same configuration.

1. Prepare the model and synthetic input while connected. Record the configuration.
2. Run a baseline task and save the result and relevant logs.
3. Disconnect the defined external network path in the test environment.
4. Run a fresh task using the synthetic input, including the retrieval or tool steps you actually need.
5. Record successes, failed dependencies, attempted connections and saved output locations.
6. Restore the original network state and classify each finding without extending it beyond the test.

If generation succeeds but document retrieval fails, the test supports local generation and exposes a retrieval dependency. It does not prove that the model itself is remote. If the entire task succeeds, the evidence supports that task under that configuration and network boundary.

An offline test cannot by itself establish what the software retained, whether it will upload queued information after reconnection or whether a future update changes its behaviour. Those questions require configuration review, logs and the relevant security assessment. Avoid calling a brief disconnect test a privacy audit.

For AI implementation planning, decide whether each remote dependency is acceptable, replaceable or a reason to change the workflow. The VRAM reference can then help size the local model rather than being asked to solve an architectural dependency.

Download the reference table (CSV). The download contains the rows shown above, with their scope and review date. It does not contain campaign results or a completed assessment of your business.

If the task uses agents and local tools, compare the boundaries with the agent sandbox reference. Local execution and isolation are related decisions, but they are not the same property.