Ollama 0.40.4 API called from PowerShell 5.1, first request takes about 30 seconds

Ollama is known as a convenient way to run local large language models, offering both a command-line interface and a graphical interface. This article focuses on its HTTP API, showing that no additional libraries are needed when using PowerShell's built-in Invoke-RestMethod. The author tested both the native /api endpoints and the OpenAI-compatible /v1 endpoints, confirming that the latter accepts familiar field names such as max_tokens and messages[].role. That compatibility could let developers reuse existing OpenAI client code when building tools for Ollama.

The test environment was Windows 11 with PowerShell 5.1 and Ollama 0.40.4, using the small qwen2.5:0.5b model. The author measured a first request duration of about 30 seconds, which includes the time to load the model into memory. This is a practical detail: even a small model incurs a noticeable startup cost, so client-side timeouts should not be set too aggressively.

Beyond timing, several pitfalls were encountered during the test: specifying charset=utf-8 in the Content-Type header to avoid encoding issues, using ConvertTo-Json with a sufficient -Depth value for nested arrays, and explicitly setting stream = $false when a single JSON response is desired instead of streaming. These are concrete, reproducible observations rather than general advice, and they illustrate the kind of adjustments needed when scripting against Ollama's API from PowerShell.