Temperature
Also known as: sampling temperature.
In plain English
At each step an LLM produces a probability for every possible next token, and temperature rescales those probabilities before one is sampled. At temperature 0 the model is nearly deterministic: it takes the top token every time, so the same prompt gives near-identical answers — ideal for extraction, classification, code, and anything factual. Around 0.7-1.0 you get natural, varied prose — the sweet spot for writing and brainstorming. Push toward 1.5-2.0 and output becomes surprising, sometimes incoherent. Temperature doesn't make a model 'smarter' or 'more truthful'; it only trades consistency for diversity. It's often tuned alongside top-p, which limits which tokens are eligible; temperature controls how evenly the eligible ones are weighted. A common mistake is cranking temperature up to fix boring output when the real fix is a better prompt — high temperature adds randomness, not quality.
Example
A developer asks a model to extract dates from invoices. At temperature 0.9 the model occasionally reformats or 'guesses' ambiguous dates differently across runs. Dropping to temperature 0 makes every run identical and predictable — exactly what a data pipeline needs. For the marketing team writing tagline options, the opposite works: temperature 1.0 produces ten genuinely different ideas instead of ten rephrasings of the same one.
Temperature vs Top-p
Temperature and top-p are two knobs on the same sampling step. Temperature reshapes the whole probability distribution — how evenly likely and unlikely tokens are weighted. Top-p (nucleus sampling) instead trims the list to the smallest set of tokens whose probabilities add up to p, then samples from just those. Temperature controls how randomly you pick; top-p controls how large the candidate pool is. Most teams adjust one, not both.
Temperature in Vincony
Vincony picks sensible sampling defaults per task so you don't have to think about temperature — factual tools run cold and deterministic, creative generation runs warmer. Each model page notes its behavior so you can choose the right tool for the job.
Browse the model catalogTry it — 750+ distinct models across 80+ providers on one account
Vincony bundles GPT-5, Claude, Gemini, Perplexity Sonar Pro, DeepSeek, Mistral, and 750+ other models on one $0/month account.