# Install Ollama. Verify the first request.

Revision 2 | 2026-10-17 | source_reported | pending_reproduction

SOURCE-BASED GUIDE. No installation or independent model benchmark performed.

Choose the native installer for your operating system, record its version, and test the API on 127.0.0.1. A successful response proves a request completed; it does not prove GPU offload or model quality.

Reference: Ollama 0.40.5 · source reference

## Before you install

Reserve a separate test folder and note the OS, CPU architecture, available memory and free disk space. Select a model you are permitted to use and budget for its download as well as runtime memory. This guide does not select a model for your hardware. The release reference is 0.40.5; all three platform paths are source guides awaiting reproduction. Keep your previous installer and configuration for recovery.

## Choose a platform path

Windows and macOS: open the versioned release source below, choose the matching native asset, and record the installed version. Do not silently replace it with an unpinned latest installer. Linux: the official installer accepts a version selector; download the script to a file, inspect it, then choose whether to run it. Installation changes your machine and may require administrator access. None of these steps is executed by this website.

## Pin the Linux install

These commands are for an isolated Linux test host. They are not a Windows PowerShell recipe. Record the installer checksum alongside the release and architecture; a checksum identifies your file but is not by itself a publisher signature.

```sh
curl -fL https://ollama.com/install.sh -o ollama-install.sh
sha256sum ollama-install.sh
# Inspect the downloaded script before running it.
OLLAMA_VERSION=0.40.5 sh ollama-install.sh
ollama --version
```

## Check the local endpoint

With the app or service running, list the installed models. An empty models array is a valid response when no weights have been installed. Select an approved local model separately; pulling a model downloads weights. Record the exact name and digest returned by the API. Never interpret a tag alone as immutable identity.

```sh
curl --fail http://127.0.0.1:11434/api/tags
ollama list
```

## Make one small request

Replace YOUR_LOCAL_MODEL with the exact name of an already installed model. On Windows use curl.exe rather than relying on a shell alias. Expect a response object containing generated text and a completion state; the exact text is nondeterministic. Save the version, model digest, options and request outcome. This is a public synthetic prompt, not a performance experiment.

```sh
curl --fail http://127.0.0.1:11434/api/generate \
  -H 'Content-Type: application/json' \
  -d '{"model":"YOUR_LOCAL_MODEL","prompt":"Return the word READY.","stream":false,"options":{"num_ctx":2048,"num_predict":16}}'
ollama ps
```

## What the first response can tell you

The model-list response identifies installed names and digests. The generate endpoint’s duration and token-count fields describe that request; timing includes the phases defined by the API, and its nanosecond durations must be converted before calculating a rate. A 16-token installation check is too short to support a stable performance comparison. Retain the response and server version before changing context or sampling.

## If the check fails

Connection refused: confirm that the intended app or service is running and that another process does not own the port. Unknown model: compare the exact name with the installed-model list. A load failure or HTTP 500 needs the error body and a sanitized log excerpt; use the troubleshooting guide. A completed CPU request is not evidence that GPU acceleration works.

## Stop, recover, and record

Use ollama stop with your selected model name to unload that model, then quit the desktop app or stop only the test service you own. Retain the previous configuration and model directory. For a version rollback, stop the service, restore the previous installer and settings, then repeat the same request. Avoid deleting the model store as a first troubleshooting step. A later reproduction record must include the binary or installer digest, not merely the release tag.

## Self-check (not site evidence)

- [ ] Installer matches OS and architecture
- [ ] Version and installer hash recorded
- [ ] Model digest and license recorded
- [ ] Loopback request returns a response
- [ ] Processor placement inspected separately

## Sources

- [Ollama v0.40.5 release](https://github.com/ollama/ollama/releases/tag/v0.40.5) | checked 2026-10-17
- [Ollama Linux installation](https://docs.ollama.com/linux) | checked 2026-10-17
- [Ollama macOS installation](https://docs.ollama.com/macos) | checked 2026-10-17
- [Ollama Windows installation](https://docs.ollama.com/windows) | checked 2026-10-17
- [Ollama list models API](https://docs.ollama.com/api/tags) | checked 2026-10-17
- [Ollama generate API](https://docs.ollama.com/api/generate) | checked 2026-10-17