Ollama
A local model runtime that provides a command-line workflow, model management and an HTTP API for running supported language and multimodal models on a personal computer or server.
What Ollama does
Ollama runs supported language and multimodal models as a local service. The command line can download, list, start and remove models, while the HTTP API lets desktop clients, scripts and self-hosted interfaces send requests. It is a model runtime rather than a complete knowledge-base, account or team-management product.
Data, network and model boundaries
Model files, caches, prompts and generated output can use substantial storage. A local runtime does not automatically make every request private: model downloads, update checks, cloud connectors and a client application may create separate network connections. Check the model license and the service binding before using real documents.
Maintenance note
This page reviews Ollama 0.32.5, model management, the command-line and HTTP API workflows, service binding, storage planning and the separate licensing of models. Content review date: 2026-08-23.
Save to your cloud drive
Open the cloud drive to get the file directly, or save it for convenient access on another device.
Quark Cloud Drive
RecommendedSave Ollama to this cloud drive
Baidu Netdisk
Save Ollama to this cloud drive
Ollama installation, local model testing and HTTP API guide
Start with a small disposable model and a local-only request. Confirm the service, storage path and response format before changing the bind address or connecting another device.
Before you start
- Check the Windows x64, macOS or Linux x64 package and reserve disk space for models, cache and logs.
- Prepare non-sensitive test prompts and a model that can be removed after the first run.
- If other devices will connect, decide the network scope, access control, proxy and backup plan first.
Installation steps
- 01
Install the matching 0.32.5 package
Extract the Windows ZIP, install the macOS DMG or unpack the Linux archive according to the platform instructions, then verify ollama --version.
- 02
Check the local service
Confirm the service is running and use the local terminal to list models or display help. Keep the default local binding during the first test.
- 03
Pull one test model
Select a model and quantization that fit the available memory and disk, record its name, size and license, and wait for the pull to finish.
Quick start
- 01
Run a local prompt
Use ollama run with a short test prompt, note first-token delay, generation speed and memory use, then stop and restart the model once.
- 02
Call the local HTTP API
Send one minimal request from the same computer, verify the model name, response format and error handling, and keep personal data out of logs.
- 03
Connect a client last
Fix the model name and API address in the client, verify the local request first, and recheck authentication and firewall rules before enabling LAN access.
Usage tips
- Inference speed depends on quantization, context length, memory, CPU/GPU and concurrency; one benchmark does not describe every model.
- An API port is not an authentication layer. Treat a changed bind address, container mapping or reverse proxy as a new access boundary.
- Model files can have separate commercial, redistribution and training-data terms; keep those records beside the model name and version.
Troubleshooting and uninstall
What should I do when a model runs out of memory?
Reduce concurrency and context, choose a smaller or lower-quantization model, inspect background processes and confirm free disk space before pulling another version.
Why can a LAN client not connect?
Check the listening address, port mapping, firewall and proxy in that order, then confirm a local request works before opening a controlled network path.
- Record models and service settingsSave the model list, client configuration and custom service parameters, and revoke access granted to clients that will no longer use the runtime.
- Stop the service and clean dataStop Ollama, then remove the application, models and cache only after checking that no client, backup or scheduled job depends on those paths.
Frequently asked questions
Is Ollama a complete chat workspace?
Its core is a local model runtime, command-line tool and API. Knowledge bases, accounts, team permissions and a full web workspace usually come from a separate client.
Does a local model guarantee that no data leaves the device?
No. Model downloads, update checks, external clients and optional connectors can use the network. Check each component and keep the first test on non-sensitive data.
Should the Ollama API be exposed directly to the internet?
Treat the API as an internal service. Use network controls, authentication at the proxy and rate limits before allowing access from outside a trusted network.