Janitor AI Generation Settings (2026): Temperature, Max Tokens and Context Size Explained
Janitor AI’s generation settings are the handful of controls that decide how a reply is produced: temperature (how predictable the wording is), max tokens (how long a reply may be) and context size (how much of the chat the model reads each turn). If you use an outside model through a proxy, there are four more under “Advanced settings”.
Most advice on these settings is a list of numbers from someone’s preset. This page does something different. We opened the settings panel on 3 October 2026 and wrote down what is actually there: every slider, its real range and what Janitor’s own text says about it. Then we matched each setting to what Janitor and the model makers publish. Where a number is somebody’s recommendation, we say whose.
Two facts change most of the usual advice, so here they are first.
- The panel was redesigned on 19 July 2026. Older guides and videos describe a “Generation Settings” window that no longer looks like that on the web. If you cannot find a setting where a guide says it is, this is why.
- On several popular models the temperature slider is not connected to anything, or is not supposed to be touched. DeepSeek, Gemini and Claude each have their own rule. A temperature you copied from a preset may do nothing, or may make every message fail.
This page is about the settings. If your question is which model to use, that is a different page: the best model for Janitor AI.
The Short Answer
| Setting | What to do | Why |
|---|---|---|
| Temperature, on JLLM | Start at 0.6, move in steps of 0.1, stay under 1.0 | Janitor’s own guide |
| Temperature, on a proxy model | Follow the model maker, not a preset | DeepSeek, Gemini and Claude all handle it differently |
| Max tokens | Leave at 0 | 0 lets the model choose; a low number cuts replies off |
| Context size, on JLLM | Nothing to set | Fixed by your plan: about 9,000 tokens free, five times more on janitor+ |
| Context size, on a proxy | The smallest size that works, often 16,000 to 32,000 | On a paid model you pay for all of it on every message |
| Top K, Top P, penalties | Leave alone | They only appear for outside models, and most players never need them |
| After any change | Press Save | Nothing applies until you do |
Where the Settings Are Now
On the web, open a chat, click the menu at the top right and choose Settings. A panel opens with three tabs: Settings, Memory and Customize.
What you see on the Settings tab depends on the Provider you pick at the top. On 3 October 2026 the list had five entries: janitor, janitor+ Router, Proxy, OpenAI and Claude.
Janitor announced the redesign on 19 July 2026. In its words, the old page “was basically one giant form that you had to scroll through forever”. The changes that matter here:
- Sections. Provider and response length are together; custom prompts, prefills and forbidden words are grouped under Instructions.
- Generation controls. For Janitor’s own model, temperature and max tokens sit together in a small section called Generation controls, closed until you open it.
- Proxy pages keep everything open. Janitor wrote that for proxy users it “left more of the advanced stuff (temperature, max tokens, etc) exposed right on the page, since thats where you guys actually tweak things”.
- A real Save button. The button “only lights up when something actually changed”, and Janitor warns you if you try to close the page with unsaved changes.
- A reset button that restores the default values.
That Save button is worth one more sentence, because it explains a lot of “I changed it and nothing happened”. Moving a slider is not enough any more. Nothing is applied until you press Save.
On the mobile app the same controls live under one menu. Janitor’s changelog for 21 May 2026 says API Settings, Generation Settings and Generation Rules “are now unified under a single ‘Model Settings’ menu on mobile”. The app itself is covered in our Janitor AI app guide.
Every Setting, With Its Real Range
This is what the web panel showed on our account on 3 October 2026. The ranges are read from the sliders themselves.
With the provider set to janitor (JLLM)
| Setting | Range | Step | Shown on our account | Janitor’s own description |
|---|---|---|---|---|
| Response length | Standard or Shorter | — | Standard | — |
| Temperature | 0 to 2 | 0.05 | 1 | ”Controls how varied replies are. Lower is more focused; higher is more creative.” |
| Max tokens | 0 to 5,000 | 50 | 0 | ”Limits how long replies can be. 0 lets the model choose the length.” |
Under the two sliders the panel shows one more line: “context is 9001 tokens · janitor+ is 5x”. There is no context slider for Janitor’s own model. The number is set by your plan.
With the provider set to Proxy or Claude
The same temperature and max tokens sliders appear, plus these.
| Setting | Range | Step | Shown on our account | Janitor’s own description |
|---|---|---|---|---|
| Context size | 0 to 128,000 | 256 | 49,920 | ”Controls how much of the conversation the AI can use at once. Lower it if replies fail to generate.” |
| Top K | 0 to 100 | 1 | 0 | — |
| Top P | 0 to 1 | 0.01 | 0 | — |
| Rep. penalty | 0 to 2 | 0.01 | 0 | — |
| Freq. penalty | −2 to 2 | 0.01 | 0 | — |
The last four sit under a heading called Advanced settings.
Three notes on reading these tables honestly.
- The “shown on our account” column is not an official list of defaults. Janitor does not publish its defaults. These are the values the panel displayed on a free account with no proxy configured. The Reset button will show you yours.
- We opened Proxy and Claude. We did not open the OpenAI provider, and we could not open the janitor+ Router, which needs a subscription. Janitor’s help page for the Router says it includes “custom prompts and advanced generation settings”.
- Janitor does not say what 0 means for the four advanced settings. A Top P of exactly 0 or a repetition penalty of exactly 0 would be strange values to send to a model. Our reading is that 0 means “off, use the model’s own default”. That is our reading, not Janitor’s statement.
Temperature: What It Does
Each time a model writes a word, it has a list of possible next words, each with a probability. Temperature changes how it chooses from that list.
- Low temperature: it almost always takes the most likely word. Replies are consistent and on-topic, and can feel stiff or repetitive.
- High temperature: it takes less likely words more often. Replies are more surprising, and more likely to forget details or stop making sense.
Janitor’s help centre has a full article on this, updated in August 2026. It splits the scale into three bands:
| Band | Janitor’s range | Janitor says it is good for | Janitor’s warning |
|---|---|---|---|
| Low | 0.2 to 0.4 | Staying on topic, keeping characters consistent | ”Can feel robotic or repetitive” |
| Medium | 0.5 to 0.7 | Natural conversation, roleplay with some spontaneity | Sometimes wanders off topic |
| High | 0.8 to 1.2 and above | Wild storytelling, chaotic characters | ”Might forget details, go off-script, or say nonsense” |
And it gives a method, which is better than a number:
“Start at 0.6 … If your bot’s responses taste too bland, sprinkle in +0.1 until you get some zing. If it starts serving word salad, dial it back down.”
It adds one limit, “keep the temperature below 1.0” unless you want chaos, and one sentence that most preset lists ignore:
“different AI models are like different cuisines - what’s mild in Claude might be extra spicy in Mistral.”
There is a small tension here that you should know about. Janitor’s guide says to start at 0.6. The slider on our account sat at 1. So if you have never touched the setting, you may already be above the value Janitor’s own guide recommends. If JLLM feels random or forgetful, lowering the temperature toward 0.6 is the first thing to try, and it costs nothing.
The Best Temperature, Model by Model
“Best temperature for Janitor AI” is really several questions, one per model. The slider is the same; what happens to the number after it leaves Janitor is not.
| Model | What to set | Source |
|---|---|---|
| JLLM (Janitor’s own) | Start at 0.6, adjust by 0.1, stay under 1.0 | Janitor’s help centre |
| DeepSeek, thinking on | It does not matter; the value is ignored | DeepSeek’s API documentation |
| DeepSeek, thinking off | 1.3 for conversation, 1.5 for creative writing | DeepSeek’s API documentation |
| Gemini (current Flash models) | Leave at the default, 1.0 | Google’s Gemini API documentation |
| Claude (current models) | Leave at the default; other values return an error | Anthropic’s documentation |
| GLM, Kimi, MiniMax, MiMo, Nemotron and others | Use Janitor’s method: start moderate, change by 0.1 | We have not checked each maker’s guidance |
DeepSeek
DeepSeek is the model family Janitor’s proxy users run the most, and it has the least obvious rule.
DeepSeek’s guide says thinking mode “does not support the temperature, presence_penalty, or frequency_penalty parameters”. It does not return an error. The setting is accepted and has no effect. Thinking is on by default on DeepSeek’s own API. So for many players the temperature slider, and the frequency penalty slider too, are not connected to anything. You can move temperature from 0.7 to 1.5 and get the same kind of replies.
With thinking off, temperature works, and DeepSeek’s own table recommends 1.3 for general conversation and 1.5 for creative writing. Both are above Janitor’s “stay under 1.0” advice, which is a good example of why one number does not fit every model.
One limit on this: the rule above is from DeepSeek’s documentation for DeepSeek’s own API. Many Janitor users reach DeepSeek models through OpenRouter, where other companies host them, and a host can handle these settings differently. We explain the two modes in the best DeepSeek model for roleplay, and the setup in how to use DeepSeek on Janitor AI.
Gemini
Google’s change log for 21 July 2026 says the sampling settings temperature, top_p and top_k “are now deprecated”, and its migration guide for Gemini 3.8 Flash tells developers to remove them. For the older Gemini 3 models Google already recommended keeping temperature at 1.0 and warned that lower values could cause looping.
So for Gemini on Janitor: leave temperature at 1, leave Top K and Top P at 0, and shape the style with your prompt. More in the best Gemini model for roleplay.
Claude
Anthropic’s documentation says that on Claude Opus 4.7 and later models, setting temperature, top_p or top_k to a non-default value returns a 400 error. In plain words: the message fails.
If you connect Claude and every message fails, check these settings before anything else. Press Reset to Defaults, then Save. More in the best Claude model for roleplay.
The pattern
Three of the largest model makers have moved away from the temperature slider in the last year, each in its own way. One ignores it, one deprecated it, one rejects it. For those models the place to change how the bot writes is the text: your custom prompt, the bot’s definition and above all its example dialogue. Our bot creation guide explains why an example beats an instruction.
Max Tokens: Leave It at 0
Max tokens is the upper limit on how long one reply can be. On Janitor the slider runs from 0 to 5,000, and Janitor’s text says: “0 lets the model choose the length.”
Three things are worth knowing.
1. It is a hard stop, not a target. The model does not plan its reply to fit the number. It writes, and at the limit it is cut. Set max tokens to 150 and you do not get neat short replies. You get long replies with the end missing.
2. If you want shorter replies, use the tools made for that. On Janitor’s own model, set Response length to Shorter. Janitor has a version of its model, jllm-short, built to write shorter, faster-paced replies; it previewed it for janitor+ subscribers on 27 June 2026 and it has been free for everyone since 15 July 2026. The redesign replaced the old “short responses” toggle with this control. On any model, the bot’s first message matters more than a slider: a long opening teaches long replies.
3. On a thinking model, a low max tokens can give you an empty reply. Many current models reason before they answer, and the reasoning counts as output. OpenRouter’s documentation says so directly: for reasoning models the limit “covers reasoning tokens and visible output together on most providers”, and a small limit “can be consumed entirely by reasoning”, which returns an empty message. If a proxy model sends back nothing, and your max tokens is a small number above 0, set it back to 0 and try again.
A rough guide to the scale, if you do want a limit: Janitor’s help centre gives 1,000 tokens as roughly 750 words.
There is a separate limit on what you write. Janitor’s help centre says a single message can be at most 10,240 characters; above that you get an error and must shorten it.
Context Size: The Setting That Costs Money
Context size is how much text the model reads on each turn: the bot’s definition, your persona, chat memory and as much of the recent conversation as fits. When the chat grows past the limit, the oldest messages are left out. That is what “the bot forgot” means.
On JLLM you cannot change it
With the provider set to janitor, there is no context slider. The panel shows the number: 9,001 tokens on our free account, with a note that janitor+ is five times that. Janitor’s subscription page says the same thing in words: the free tier has a “context limit of around 9k tokens”, and janitor+ gives “5x more context”.
So on Janitor’s own model, the only two ways to make the bot remember more are to keep the permanent parts of the bot small, and to subscribe. Janitor’s token guide advises keeping the permanent part under about 1,500 tokens. We cover this in Janitor AI chat memory.
On a proxy, the slider stops at 128,000
With an outside model, the context slider runs from 0 to 128,000 tokens in steps of 256. On our account it sat at 49,920.
This is worth a moment. Many current models advertise a context window of 1 million tokens. Janitor’s slider cannot send more than 128,000 of them. A 1M window on the model’s price list does not mean a 1M chat on Janitor. We are not complaining about this; as the next table shows, almost nobody should send that much anyway.
Why smaller is usually better
Every turn sends the whole context again, and on a paid model you pay for all of it, every time. Here is one example, using DeepSeek V4.1 Flash at the price OpenRouter listed on 2 October 2026, $0.30 per million input tokens.
| Context size | Input cost of one turn | 200 turns |
|---|---|---|
| 8,192 | $0.0025 | $0.49 |
| 16,384 | $0.0049 | $0.98 |
| 32,000 | $0.0096 | $1.92 |
| 49,920 (the value on our account) | $0.0150 | $3.00 |
| 128,000 (the maximum) | $0.0384 | $7.68 |
These numbers are our arithmetic and they are a worst case. They assume the context is completely full on every turn, with no caching discount. Output is extra, but small: 200 replies of 500 tokens add about $0.12 on this model.
The reading is simple. On a cheap model the difference is a few dollars. On an expensive one, multiply. The same 49,920 tokens on a model that costs $2 per million input tokens is $20 for 200 turns.
There are two more reasons to keep it small.
- Replies can fail when the context is larger than the model accepts. That is what Janitor’s own hint means: “Lower it if replies fail to generate.” Some models on a proxy have a window smaller than 49,920 tokens. If a model returns errors from the first message, lower the context size before you change anything else.
- Free models have request limits, not token limits, so context size does not cost money there. It still slows the reply.
A practical way to choose. Start between 16,000 and 32,000. Keep chat memory up to date, so the important facts are always in the context. Raise the size only when the bot forgets things that memory should not have to carry.
The Advanced Settings: Top K, Top P and the Penalties
These four only appear for outside models. Janitor’s panel gives them no description, so here is what each does, from OpenRouter’s parameter documentation.
| Setting | What it does | The usual default |
|---|---|---|
| Top K | Limits the choice to the K most likely next words. A value of 1 always picks the most likely one. | 0, which means off |
| Top P | Limits the choice to the most likely words whose probabilities add up to P. Lower is more predictable. | 1.0, the full range |
| Rep. penalty (repetition penalty) | Makes the model less likely to reuse words already in the text. Too high makes the writing less coherent. | 1.0, no penalty |
| Freq. penalty (frequency penalty) | Lowers the chance of a word the more often it has already appeared. Negative values encourage reuse. | 0 |
OpenRouter also explains something useful about leaving settings alone: when a setting is absent from a request, it does not fill in a value of its own, “so the provider applies its own default”.
Our advice is to leave all four where they are unless a specific problem sends you there.
- They overlap with temperature. Temperature, Top K and Top P all change the same choice. Moving three of them at once makes it impossible to tell which one helped.
- Not every model accepts them. DeepSeek ignores the frequency penalty in thinking mode. Gemini has deprecated Top K and Top P. Claude returns an error for a non-default Top K or Top P.
- A preset’s numbers belong to the preset’s model. A list of four values tuned for one model is not a setting for another.
If you do want to try one, change one, send several messages, and compare.
What about repetition on JLLM?
People search for a repetition penalty for Janitor’s own model. There is none in the panel. The advanced settings are only for outside models.
Janitor has been open about the cause. Its changelog for 30 April 2026 says JLLM getting stuck repeating itself “is a known issue at the model level so no amount of tuning fully solves it”. On 18 July 2026 it reported an update to the model that reduced detected loops “from 6/64 generations to 2/64” in its own tests. So on JLLM the fix for a loop is not a slider. Reroll the reply, or edit the repeated phrase out of the last message so the model stops copying it. Our guide to why AI roleplay repeats itself covers the habit in general.
The Three Settings Next Door: Custom Prompt, Prefill and Forbidden Words
These are under Instructions in the same panel. They are not sliders, but they do more for the style of a reply than any of the sliders do.
Custom prompt. Your own standing instructions to the model. Since the July redesign you can save several and switch between them. On 20 July 2026 Janitor raised the size limit for a saved custom prompt to 500,000 characters, from 10,000. That is a limit, not advice: a custom prompt is sent on every turn and takes room from the conversation, like any permanent text.
Prefill. Janitor’s panel describes it this way: it “puts words in the AI’s mouth: your text is slipped in as the start of its reply, right after your latest message, and it continues in that voice and format. Applies to every reply while on.” It is a strong tool for fixing format. Note that Google’s migration guide for Gemini 3.8 Flash tells developers to remove prefilled model turns, so do not assume prefill works on every model.
Forbidden words. A list of up to 10 entries of up to 50 characters each. Janitor is clear about what it is: the list is “sent as a note with your latest message telling the AI to avoid these words and phrases. It steers strongly but is best-effort — not a hard filter.” So it is an instruction, not a block. A forbidden word can still appear.
None of these three changes how text looks. Italics, bold, text colors, fonts and text streaming are display settings, covered in Janitor AI text formatting and commands.
When a Setting Seems to Do Nothing
| What you see | Likely cause | What to do |
|---|---|---|
| You changed a slider and replies are the same | You did not press Save | Press Save; the button lights up when something changed |
| Temperature changes nothing on DeepSeek | Thinking is on, and temperature is ignored | Turn thinking off if your provider allows it, or change the prompt |
| Every message fails on Claude | A non-default temperature, Top K or Top P | Reset to Defaults, then Save |
| Replies stop in the middle of a sentence | Max tokens is set too low | Set max tokens to 0 |
| Empty replies from a proxy model | A thinking model used the whole max tokens limit | Set max tokens to 0 |
| Errors from the first message on a proxy | Context size is larger than the model accepts | Lower the context size |
| The bot forgets early events | The chat is longer than the context | Update chat memory; on a proxy, raise context size |
| You changed the proxy URL and nothing changed | The page needs a reload | Janitor’s panel says to refresh after changing a proxy URL |
| The setting from a guide is not there | The guide is older than 19 July 2026, or you are on JLLM | Advanced settings only appear for outside models |
| The model’s reasoning shows in the chat | The model thinks before it answers | See below |
On that last row: Janitor’s changelog for 23 January 2026 announced a “Show Thinking” toggle in Generation Settings, switched on by default. We did not see that toggle in the redesigned web panel on 3 October 2026 for the providers we opened, so we cannot tell you where it is now. If reasoning text bothers you, the dependable fix is a model where thinking is optional.
If a reply fails for a reason that is not a setting, Janitor AI not working lists the errors and what each one means.
Our Recommendation
- On JLLM, change one thing: temperature. If replies feel random, go down toward 0.6. If they feel flat, go up by 0.1 at a time. Use Response length for shorter replies. Leave max tokens at 0.
- On a proxy, set context size on purpose. It is the only setting here that changes your bill. Start between 16,000 and 32,000.
- Before you set a temperature on a proxy model, find out whether the model listens to it. DeepSeek with thinking on does not. Gemini and Claude want the default.
- Leave the four advanced settings alone until a specific problem gives you a reason.
- Do not copy a preset’s numbers to a different model.
- Press Save.
- For style, write instead of sliding. A custom prompt, a prefill and good example dialogue change how a bot writes far more than 0.1 on a slider.
Frequently Asked Questions
What is the best temperature for Janitor AI?
There is no single best number, because it depends on the model. For Janitor’s own model, JLLM, Janitor’s guide says to start at 0.6, add 0.1 if replies are bland, lower it if replies stop making sense, and stay below 1.0. For a proxy model, follow the model maker: DeepSeek ignores temperature while thinking is on and suggests 1.3 to 1.5 with thinking off, Google tells you to leave Gemini at the default of 1.0, and current Claude models return an error for any value except the default.
What does temperature do on Janitor AI?
Temperature controls how predictable the next word is. At a low value the model almost always picks the most likely word, so replies are consistent and can feel stiff. At a high value it picks less likely words more often, so replies are more surprising and can lose the plot. On Janitor the slider runs from 0 to 2 in steps of 0.05.
What should max tokens be on Janitor AI?
Leave it at 0 unless you have a reason. On Janitor, 0 means the model chooses the length of its reply. The slider goes up to 5,000 in steps of 50. A number above 0 is a hard stop, not a target, so a low value cuts replies off in the middle of a sentence. On a model that thinks before it answers, a low value can produce an empty reply, because the thinking uses up the limit first.
What is the best context size for Janitor AI?
On Janitor’s own model you cannot set it: the free plan shows a fixed context of about 9,000 tokens, and janitor+ gives five times more. On a proxy the slider runs from 0 to 128,000 tokens. Use the smallest size that keeps your story working, because on a paid model you pay for the whole context again on every message. Around 16,000 to 32,000 tokens is enough for most chats if you keep chat memory up to date.
What are the best generation settings for DeepSeek on Janitor AI?
Leave max tokens at 0 and set context size to what your chat needs. For temperature, it depends on thinking mode. DeepSeek’s documentation says that with thinking on, which is the default on its own API, the temperature setting is accepted but has no effect. With thinking off, DeepSeek’s own table suggests 1.3 for general conversation and 1.5 for creative writing. That guidance is for DeepSeek’s own API; other hosts of DeepSeek models may behave differently.
Why are there no advanced generation settings on JLLM?
Janitor only shows the advanced settings when the provider is an outside model. With the provider set to janitor, the panel has response length, temperature and max tokens, and nothing else. With a proxy or a Claude key, it adds context size, Top K, Top P, repetition penalty and frequency penalty.
How do I reset Janitor AI generation settings?
Open the chat menu, choose Settings, and use the Reset button in the generation controls. On a proxy it is labelled Reset to Defaults. Then press Save. Since the redesign of 19 July 2026 nothing is applied until you press Save, and Janitor warns you if you close the panel with unsaved changes.
Why does changing the temperature on Janitor AI do nothing?
Three common reasons. You moved the slider and did not press Save. You are on a DeepSeek model with thinking on, which ignores temperature without giving an error. Or you are on a current Gemini model, where Google has deprecated the temperature setting. In each case, change the style through your prompt or the bot’s example dialogue instead.
The slider ranges, steps, displayed values, the provider list and the quoted descriptions of each control were read from Janitor AI’s web settings panel on a free account on 3 October 2026. The redesign, the mobile menu, the custom prompt limit, the shorter-reply model and the repetition figures are from Janitor’s newsroom and changelog. The temperature bands and the 0.6 starting point are from Janitor’s help centre, as are the context limits by plan and the message length limit. The DeepSeek, Gemini and Claude rules are from each company’s own API documentation. The descriptions of Top K, Top P and the penalties, and the note on reasoning and max tokens, are from OpenRouter’s documentation. Every cost figure is our own arithmetic on a listed price. We have not run a controlled test of how each setting changes the writing of each model.
Product names and trademarks are the property of their respective owners. Arcanum is not affiliated with Janitor AI, OpenRouter, DeepSeek, Google, Anthropic or any model maker named here.