How to make a local AI model give the same answer every time

Ask a language model to score a CV from 1 to 10 and it will happily give you a number with a reason. Ask again, and the number can change. That is fine for brainstorming, but not for anything you compare: if two candidates are scored on different runs, a one-point gap might just be luck. The fix is one setting, temperature, and the quickest way to see what it does is to ask the same question several times side by side. Below is how to do that in the osFoundry desktop app with a model running on your own computer, and then how to make the answers line up.

1. Open three panels with the same model

In the app, go to Models, then Playground. The Playground sends one prompt to several panels at once, which is usually used to compare different models. Here every panel gets the same one. Click Add on the right edge until there are three panels, then open each panel's model list and pick the same model in all three. In the screenshots it is Qwen2.5-7B-Instruct, a local model marked "local", so the prompt never leaves the laptop.

Three panels is not a magic number. It is simply enough to see a disagreement in one click, instead of pressing Run, reading, and pressing Run again.

2. Paste the CV and ask for a score

Paste the CV into the prompt box and add the question at the end, for example: "Score this CV from 1 to 10 for a junior data analyst role. Start with the score, then one reason." Asking for the score first keeps the answers short and easy to compare. Then click Run.

Three panels, one model, one prompt, and the scores come back as 6, 8 and 7. Nothing is broken. For every next word, the model gives each possible word a probability, and the app then draws one of them at random, weighted by those probabilities. If "7" is the most likely score but "8" and "6" still have a real chance, some runs will draw them. The reason after the score follows whichever number was drawn, which is why each answer sounds convincing on its own.

3. Set Temp to 0

Click the settings icon at the top right of the Playground. A row appears above the prompt with Temp, Max Tokens and the system prompt. Set Temp to 0.

Temperature decides how much chance the less likely words keep. Higher values flatten the probabilities so that unlikely words get picked more often, which helps with creative writing. Lower values sharpen them toward the favourite. At 0 there is no draw at all: the model takes the most likely word every time, so the same prompt gives the same text.

4. Run it again

Click Run with the same CV and question.

All three panels now answer 7, word for word. That is the point of the setting: the same input gives the same output, so a difference between two CVs comes from the CVs, not from the draw.

It doesn't make the score right. A temperature of 0 picks the model's single most likely answer, and that answer can still be wrong, or can move when you change the wording of the question. Read the reason it gives, keep the question identical for every CV you compare, and treat the number as a way to sort, not a verdict. For scoring, sorting and grading, Temp 0 is the sensible default; for writing a cover letter or brainstorming names, leave it higher so you get variety.

The temperature glossary entry explains the setting in more detail, and the app is free from the osFoundry download page for macOS, Windows and Linux.