Published November 20, 2026
Best Janitor AI Settings for Better Replies _ Temperature and Tokens
Most "Janitor AI gives bad replies" complaints are not about the model. They are about two sliders and a token budget. JanitorLLM ships with defaults that work against you, and the fix Janitor's own help centre recommends takes about four seconds. Here is what to change, what to leave alone, and why.
Start with temperature
Temperature controls how varied and unpredictable the wording gets. Low values make replies focused and safe. High values make them wild, then incoherent. On JanitorLLM the default is 1. Janitor's own guidance says to start at 0.6, adjust in steps of 0.1, and stay under 1.0. Read that again: the shipped default sits above the ceiling the documentation recommends.
So do this first. Open a chat, open the top-right menu, go to Settings, and find temperature under Generation controls. Set it to 0.6. Chat for a few exchanges. Flat or stiff? Add 0.1. Repetitive or loopy? The July 2026 JLLM update already cut looping a lot, but a high temperature brings it back, so drop 0.1 instead. One change at a time, or you will never know what helped.
On a proxy model, follow the model maker's advice rather than a universal number. DeepSeek, Gemini, and Claude each handle temperature differently, and on some models the slider is effectively ignored. The setup walkthrough on our homepage shows where these controls live after the July 2026 redesign.
Max tokens: leave it at 0
Max tokens caps how long one reply can be. On JLLM it ships at 0, which means unlimited: the model decides. Leave it there. The classic beginner mistake is setting a low number to get shorter replies, then wondering why every answer cuts off mid-sentence. If replies run long, use the Response length setting instead and switch Standard to Shorter. That is what it is for.
The real limit: your 9,000-token budget
Here is the part most setting guides get wrong. On the free plan, JLLM reads roughly 9,000 tokens of context at once, and you cannot change that number. There is no context slider on JLLM. Guides telling you to set 6,144 or 8,192 are describing a field that does not exist when Janitor's own model is selected.
That budget is shared. The character's permanent definition eats first, then your persona, then the chat history. A 3,000-token character card leaves you about two-thirds of the budget for the actual story, and the bot starts forgetting early. This is why the same settings feel great on one character and broken on another. If a bot forgets your name by message twenty, slim the card or your persona before you touch a slider. janitor+ raises the budget to about five times the free amount, which is the real upgrade in that tier.
On a proxy, context size is a real setting and a real cost, because you pay for the whole context on every message. Around 16,384 is a sane default. Bigger is not smarter. It is slower, pricier, and past a point the model gets worse at using it.
Your persona is a setting too
Your persona sits in your Profile under the Persona tab, and it is part of every prompt. Keep it to 50 to 100 words: who you are in the story, how you look, how you talk. Long personas burn budget on every single message. Short ones get remembered. Keep a few personas for different genres and swap them per chat instead of writing one giant one.
The Instructions section holds two more quiet levers. A short custom prompt can set the writing style you want, in plain words. Forbidden words can kill a catchphrase the bot will not drop. Use both sparingly; a page of instructions fights the character card and eats the same budget.
A simple routine that works
Set temperature to 0.6. Leave max tokens at 0. Trim your persona to under 100 words. Pick characters with lean cards. Then chat, and adjust temperature by 0.1 steps only when something specific bothers you. If replies still loop or come back empty after that, the problem is probably not settings at all, and our API error fixes will take you further. Settings are seasoning. The budget is the meal.