Does Ollama run locally or in the cloud?

Ollama supports both. A downloaded local model uses your computer for inference. A cloud model sends the request to hosted compute, even when you open it through the Ollama desktop app or a local server. Check the selected model and connection before entering private material. Ollama API introduction

Documentation checked: 12 October 2026. These instructions were checked against official documentation and source code; we have not executed this setup or audited its network traffic.

Why a desktop app can still use cloud AI

The chat window and the model can run in different places. Ollama's cloud documentation currently uses gemma4:cloud for its app or CLI, but gemma4:31b for a direct request to the hosted API. Cloud models do not require their weights to be downloaded. A model's name needs to be interpreted alongside the destination. Ollama Cloud

An integration pointing at http://localhost:11434 connects to an Ollama server on your computer. That server can still handle cloud models after sign-in. An integration pointing directly at https://ollama.com uses the hosted service. Seeing “localhost” therefore answers where the first connection goes, rather than proving where inference happens. API destinations

1. Check the exact model before sending a document

Start with Ollama installed and a harmless test message. In PowerShell or another terminal, list the models known to the server:

ollama list

ollama ls is also supported. A listing alone does not prove that every entry contains local weights: the inspected CLI can list remote models too. Copy the complete name of the model selected in your chat app, then replace the placeholder below:

ollama show "MODEL_NAME_FROM_LIST"

The inspected implementation prints Remote model and Remote URL when the model information includes a remote host. If those fields appear, investigate that destination before proceeding. Absence of those fields is not a security certification, particularly if your CLI points at a server on another machine. CLI implementation inspected

For a local model already loaded by your ordinary test chat, run:

ollama ps

Its PROCESSOR column reports CPU, GPU or split memory placement. An empty listing means there is no loaded model reported at that moment; it does not establish cloud use or privacy. Running-model reference, processor explanation

2. Disable Ollama cloud features on Windows

For the standard Windows desktop installation:

1. Quit Ollama through its system-tray menu.
2. Open Windows Settings or Control Panel and search for environment variables. Choose Edit environment variables for your account.
3. Under your user variables, create OLLAMA_NO_CLOUD with value 1.
4. Save with OK/Apply, then launch Ollama from the Start menu. Open a fresh terminal for subsequent checks. Windows environment and restart guidance

Ollama documents this variable as disabling its cloud models and web search. It also supports disable_ollama_cloud in ~/.ollama/server.json, but you do not need both approaches. Use the persistent user variable here to avoid replacing an existing configuration file. Cloud controls and Windows environment instructions

Changing a variable only inside an existing PowerShell window does not reconfigure an already running tray app. Windows applications inherit environment settings when started; quitting and reopening the app matters. A service or container needs its own environment and restart procedure. Windows restart guidance

3. Confirm the setting, then keep a short record

Press Windows + R, open %LOCALAPPDATA%\Ollama, and inspect the latest server.log. The documented confirmation is Ollama cloud disabled: true. Check a fresh startup entry rather than relying on an old line. Log locations, confirmation message

Record the app/server location, selected model, date and confirmation in a plain text note. This is your configuration record, not proof that every installed application is offline. Keep sensitive prompts and documents out of troubleshooting logs you share.

The control belongs to Ollama. It does not disable another application's separate hosted provider, document upload, web tool or cloud-synchronised folder. It is also not a firewall or an air gap. Ollama's FAQ describes internet model downloads and automatic Windows updates separately. Review the surrounding application as well as the inference server. Configuration implementation, Ollama FAQ

Free local models versus hosted access

Running downloaded models on your hardware does not require an Ollama subscription; your hardware, storage and electricity still matter. Cloud access can accommodate models your PC cannot comfortably run. The current pricing page lists starter usage credits and paid options, so check the selected model and account allowance rather than assuming unlimited free cloud use. Current plans

Ollama's March 2026 privacy policy says locally processed content stays on the machine, while cloud prompts and responses are processed transiently without model training. It also describes metadata collection and service providers. These are provider statements, not a network audit by this website. Cloud processing still sends the input away from your computer. Ollama privacy policy

If the check does not match what you expected

• The confirmation is missing: recheck the variable's spelling and value, quit/relaunch the tray app, and inspect the correct server's latest log.
• A cloud model stops working: cloud access was deliberately disabled. Select a suitable downloaded local model instead; do not re-enable cloud merely to make a private-document task run.
• The model appears local but the chat app uses a hosted provider: check that app's active provider, model and connection. Ollama settings do not control an independent provider.

For choosing a fitting local model, continue with Can my PC run local AI?. For an existing local model's speed, use Why does my local AI model run slowly?.