Model Parameters

What we're trying to control

Every model call so far has used whatever defaults ChatGoogleGenerativeAI picked for us. Two parameters are worth knowing how to control directly:

• temperature: how willing the model is to pick a less-likely next word instead of the most-likely one. Low (near 0) tends to produce the same or very similar answer every time you ask; high (near 1 or above) tends to produce more varied answers to the exact same question.
• max_output_tokens: a hard cap on how long the reply is allowed to be. A token is roughly a word-piece, not exactly a word or a character, models don't count length in letters, they count it in tokens.

steady_model = ChatGoogleGenerativeAI(model="gemini-3.5-flash-lite", temperature=0)
creative_model = ChatGoogleGenerativeAI(model="gemini-3.5-flash-lite", temperature=1.0)

An honest surprise

Run this lesson and you'll likely see a warning like:

UserWarning: Model 'gemini-3.5-flash-lite' uses fixed sampling defaults;
the sampling parameter(s) temperature will be ignored.

And sure enough, comparing the temperature=0 answers to the temperature=1.0 answers, both lists look similarly varied, temperature had no visible effect. This isn't a bug in the lesson, it's a real, useful fact: gemini-3.5-flash-lite is a small, fast, distilled model, and Google has fixed its sampling settings internally, this specific parameter simply doesn't apply to it, no matter what value you pass.

This is worth sitting with, because it generalizes: not every model supports every parameter, and the shared LangChain interface (the same ChatGoogleGenerativeAI class, the same temperature= keyword) doesn't guarantee every underlying model actually honors every setting. When something you configure doesn't seem to change anything, checking for a warning like this one, or the provider's own documentation for that specific model, is the right instinct, not assuming your code is wrong.

If you want to see temperature actually working, the same code would behave differently on a bigger model in the same family (like gemini-3.5-flash, without -lite) that doesn't fix its sampling settings, at the cost of a much lower free daily request limit, which is exactly why this project defaults to the -lite model everywhere else.

max_output_tokens: this one does work

short_model = ChatGoogleGenerativeAI(model="gemini-3.5-flash-lite", max_output_tokens=10)
long_answer = short_model.invoke("Explain what LangChain is.")

Unlike temperature, max_output_tokens is respected. Asking a naturally long question ("Explain what LangChain is.") with a tiny 10-token cap produces a reply that's visibly cut off mid-thought, something like "**LangChain** is an", and then nothing more. The model wasn't finished, it simply ran out of allowed length.