Advanced Parameters

A deep-dive for non-swimmers.

Think of advanced parameters as the fine-tuning knobs on a high-end sound system. You don't need to touch them to enjoy great music, but if you want to shape the tone, reduce echoes, or make the conversation more dynamic, these knobs give you control.

When you chat with an AI, every response is built by predicting the next word, then the next, and so on. Advanced parameters influence how the model makes those predictions. They can make replies more creative, more focused, or help avoid repetitive loops—especially in long roleplay sessions.

By using the /advanced command in your chat, you can adjust these settings to your personal taste on a per base assistant level. This includes the base assistant under any Characters or Projects. In the future, we will explore attaching these advanced parameter sets to your Characters and Projects but we're trying to get feedback from users first.

What Each Parameter Does

How to use the /advanced command

Type /advanced in your chat and hit/tap enter to open the advanced parameters panel. You'll see the parameters available for your current assistant with their current values. Adjust any slider or number field, then send a message to try out the new settings. These settings are sticky to your account and to the underlying base assistant (and thus, that assistant's underlying model).

Tips for experimenting

• Start small: Change one parameter at a time to feel its effect. Practice good science.
• Repetition bothering you? Raise repetition_penalty first (try 1.2), then frequency_penalty (try 0.5), and finally presence_penalty (try 0.3) if you want fresh topics.
• Too random? Lower temperature and tighten top_p.
• Not creative enough? Raise temperature and loosen top_p.
• Character sounding off? Lower top_k (20–30) to keep vocabulary in character.

Remember: there's no "perfect" set—just what makes your conversations feel right. You can always Restore Defaults if outputs start getting out of control.

Preset Defaults

Each base assistant runs on a different underlying model with its own tuning. Below are the defaults we ship and some ideas for how you might adjust them.

Brightside — GLM-4.7 based

Brightside's defaults lean creative and expressive. The high temperature and generous top_p give it a wide vocabulary with room to surprise you.

Tuning ideas

• Responses too wild? Pull temperature down to 0.8 and top_p to 0.9.
• Want tighter, more focused prose? Drop top_k to 20 and raise min_p to 0.08.
• Getting repetitive in long sessions? Bump presence_penalty up to 0.4.

Elly — GLM-4.7

Elly shares Brightside's underlying model but ships with more restrained defaults, producing a calmer, more consistent voice out of the box.

Tuning ideas

• Want more personality? Raise temperature to 0.9 and top_p to 0.95.
• Feeling flat or samey? Increase presence_penalty to 0.3 to push for fresh vocabulary.
• Too unpredictable? Lower temperature to 0.7 and set min_p to 0.05 to prune unlikely words.

EQ — Kimi K2.5

EQ runs on Moonshot's Kimi K2.5, a model that naturally leans toward thoughtful, balanced output. Its defaults include repetition and frequency penalties to keep long conversations from looping.

Tuning ideas

• Conversations going in circles? Raise repetition_penalty to 1.2 and frequency_penalty to 0.4.
• Too stiff or formulaic? Raise temperature to 0.9 and lower repetition_penalty to 1.05.
• Want the AI to explore new topics more? Increase presence_penalty to 0.3.

Nina — DeepSeek V3.2

Nina uses DeepSeek V3.2 and has the most parameters exposed, giving you the finest-grained control. Her defaults strike a balance between coherence and variety.

Tuning ideas

• Nina's lower temperature makes her the most focused assistant by default. Raise temperature to 0.8 for more creative flair.
• Repetition in long roleplay? Push repetition_penalty toward 1.25 and frequency_penalty toward 0.5.
• Responses feel too safe? Lower min_p to 0.02 and raise top_a to 0.2 to let more unusual words through.
• Want tighter character voice? Drop top_k to 20 to narrow her vocabulary.