AI

Local models

Run RolePort's AI for free on your own computer, with nothing leaving your machine.

Last updated 31 October 2026

A local model costs nothing and keeps your CV on your computer. It's slower than a hosted provider and needs a reasonably capable machine.

Set up Ollama

1. Install Ollama.
2. Download a model:

ollama pull gemma4:12b

3. In Settings → AI provider, choose Ollama. Your downloaded models load on their own.
4. Pick the model, leave Thinking on Off (fastest), then click Test connection and Save.

In Docker, RolePort can't see Ollama at localhost, so set the Base URL to:

http://host.docker.internal:11434/v1

Using Ollama from RolePort Lite

Ollama only answers websites you allow, so allow the Lite address once. Settings shows these commands with your address filled in.

Windows, in PowerShell, then quit Ollama from the system tray and open it again:

setx OLLAMA_ORIGINS "https://your-lite-address"

macOS, in Terminal, then quit Ollama from the menu bar and open it again:

launchctl setenv OLLAMA_ORIGINS "https://your-lite-address"

Linux, run sudo systemctl edit ollama.service, add these lines and save, then run sudo systemctl restart ollama:

[Service]
Environment="OLLAMA_ORIGINS=https://your-lite-address"

If your browser asks to let the site access devices on your local network, choose Allow. Safari can't always reach Ollama, so use Chrome, Edge or Firefox.

LM Studio and other servers

Choose Custom and enter the server's Base URL. LM Studio's default is http://localhost:1234/v1. An API key is optional. Any server that speaks the OpenAI chat format works.

Choosing a model

• Use a 12B or larger model for reliable results, such as gemma4:12b or a Qwen 14B to 32B.
• Smaller models can misjudge the fit gates and struggle with long prompts.
• Set Thinking to Off on thinking models, which is about ten times faster. Turn it up for a job you care about.

What works on small models

Every feature is built to work on a 12B model. RolePort keeps the AI's job small and checks the results in code:

• Overall scores are worked out by RolePort, not the model.
• Quotes from postings have to really appear in the posting.
• Pasted job fields have to appear word for word in the text.
• Facts on your CV are copied from your profile.
• Mock interviews send less context to local models.

How long things take

A fit evaluation takes about 30 to 60 seconds on a 12B model with a mid-range GPU, and a few seconds on a hosted provider. Tailoring a CV or planning an interview can take a few minutes. The first request after starting can take a minute while the model loads. Keep the page open while it works.