This is an unofficial guide to Moonshot AI products. I'm not affiliated with Moonshot AI.
Kimi K2 Thinking for high density Character Cards
INFO
A collab with rotten_script_79
He writes bots that are insanely detailed and well cooked! While beta-testing each other's works we stumbled over hiccups that could only be explained by this:
Detailed Character Cards + rigid System Prompt + MoE-Reasoning Monsters = Monolithic Chaos.
To fix this, we engineered a specialised System Prompt specifically for token-heavy, detail-rich roleplay settings.
Overview
Kimi-K2-Thinking is an MoE and reasoning Monster!
It is censored when it comes to underage characters. Anything else is ... wild.
When it gets a task it will fulfill it. So prompting it for roleplay is ... a challenge!
The prose is nice from the get go. Tropes and cliché are almost no issue. ALMOST. It tends to go very intense with dark topics. So I will do something I have never done before:
Warning! If you have depression, PTSD, or other problems with your mental health... Do not use Kimi-K2-Thinking!
I'm serious, love! Self care is top priority.
Generation Settings:
- Max Token Output: unlimited or at least 2K
- Temperature: 0.85 - 1.0
- Max P (Top P): 0.95 - 0.98
- Min P: 0.0
- Top K: 0 (IMPORTANT!!!!)
- Penalties: disabled... all of them
- Reasoning enabled.
Why?
Kimi-K2-Thinking is a monster. We are talking about a Mixture of Experts with Reasoning. That thing can't perform right with restrictions of any kind. The more room (output tokens) and freedom (no penalties, no restricting samplers) you give it, the better it will perform.
Pre-History / System Prompt
Post-History / Jailbreak