Why do AI roleplay replies cut off mid-sentence?
Updated
A reply that stops mid-sentence nearly always hit its length limit. In SillyTavern the default Max Response Length of 300 tokens is the usual culprit; raise it to 600 to 800. In JanitorAI raise Max new tokens. Other causes are a context so full the reply had no room, a stop string matching too early, or a dropped connection. On UnsavedAI a single reply can be up to 8,192 tokens.
Try it now, free, no account neededStart a story with The NarratorAny story you want, you choose1. The reply length setting is too low
Every app has a ceiling on how long the model may write: Max Response Length in SillyTavern, Max new tokens in JanitorAI, max_tokens in code. When the model reaches it, generation simply stops, wherever it is. SillyTavern's default of 300 tokens is roughly 220 words, shorter than many roleplay replies.
Set it to about 700 for gemma-4-12b or 800 for gemma-4-31b. That allows several full paragraphs. The model still stops earlier when it is done; the setting is a ceiling, not a target.
2. The context is nearly full
The model's limit covers the prompt and the reply together. If your app sends a prompt that nearly fills it, we shorten the reply so it fits instead of failing, and that shortened reply can end abruptly. This tends to show up only in long chats, and gets worse as they grow.
Fix it by setting the app's context size to 56,000 for our Gemma models, so there is always room left to answer. See context size for roleplay for why that number.
3. A stop string fired
Stop strings tell the model to halt when it writes a certain text, often your name followed by a colon so it does not speak for you. If a stop string is too common, such as a single newline or a word the character uses, replies end early, usually at a clean point rather than mid-word. Check custom stopping strings in SillyTavern and any preset you imported.
4. The stream was interrupted
If the connection drops or you press stop, you keep whatever arrived. This looks random: different lengths, different moments, sometimes mid-word. Refresh and try again; if it repeats, check your connection, or whether a browser extension is interfering.
Finishing a cut-off reply
You do not have to throw it away. SillyTavern's Continue option asks the model to keep writing from where it stopped, in the same message. In chats on UnsavedAI itself, the Continue button under a reply does the same. Fix the setting first, or the continuation can stop short too.
Questions people ask
Why did replies get shorter after an update?
App updates and preset imports sometimes reset reply length to a small default. Check Max Response Length or Max new tokens first.
Why does the cut-off happen only in long chats?
That points to a full context rather than the reply setting. Lower your app's context size to 56,000 so there is always room for the answer.
Is there a maximum reply length?
On UnsavedAI one reply can be up to 8,192 tokens on either plan. Roleplay rarely needs more than 1,000.
The reply is complete but too short. Same problem?
No. That is the model choosing to stop. Longer example dialogue in the card and longer messages from you usually draw longer replies.
