Evening-Truth
This is an unofficial guide to Deepseek products. I'm not affiliated with Deepseek.
Deepseek V3.2 Thinking Preset for more Realism and Smut
I gave this one an update.
The lines have been reorganized and refined. NSFW stuff should be less rushed now. Responses more on point but still immersive.
It shows once again that simplicity can be better.
Try it, see if you like it.
If you don't, the previous version is still here.
Remember to change the content rules to your preference!
INFO
Overview
Deepseek V3.2 is a damn good storyteller and world builder with impressive context coherency and creativity. It sticks to the script like a clingy octopus. Lore heavy bots or well known characters are a walk in the park for it. Character consistency is intense.
However V3.2 isn't the best in eroticism and smut. Kimi K2 and GLM 4.6 perform way better in that area.
Can handle relatively big context sizes. First major coherency hiccups came at around 20K tokens without any entries in Long Term Memory.
Creativity 9/10
Complex Characters 10/10
Context Coherency 10/10
Horror, Dead Dove 9/10
Smut, ERP, Kink 6/10
Romance 8/10
Fantasy 9/10
The annoying bits
I said it before and I will say it again... Deepseek is ADHD in corporate form. And they did a thing again... and forgot to let users know. Probably got distracted. qed.
Anyway, if you make your calls to the Deepseek API you have to adjust the temperature to at least 1.5 for creative writing. Yeah, you read that right. The thing they did is a Temperature offset that lowers Temp to 0.3 if you set it to 1.
You might also find that there is more of an avoidant behavior in responses. No hard censoring, just avoidance. That's new too.
I can hear your sigh from here and I feel you.
If you have a third party provider that runs DS models on their own hardware... you'll likely get the performance we all know and loved and a Temperature setting that doesn't gaslight you.
Just saying.
Tips:
- Deepseek thrives on clear instructions.
So if you want the Roleplay to go in a specific direction, tag it with e.g. Adventure, Romance, Smut. The same goes with things you don't want to see. Just be clear about it. Use headers like Forbidden Styles/Genres/Tones and Allowed Styles/Genres/Tones. - Use the Long Term Memory only for backstories or other details that are not already in the Chat History. It will mess up the coherency.
- Deepseek doesn't really understand the concept of internal thoughts in roleplay. So you need to give your thoughts a distinkt formatting.
SillyTavern users:
- Set the Prompt-Post-Processing to semi-strict or strict to help with coherency. It doesn't perform well when it's set to none* .
Other Platforms:
- If you don't have the option for post-history-instructions, try putting them into chat memory or at the very beginning of the prompt.
Generation Settings:
Temperature: 0.75 - 1.0 (at least 1.5 when using DS API)
Max P (Top P): 0.90 - 0.98 (I prefer 0.95)
Top K: 0 (IMPORTANT!!!!)
Penalties: (using -2 to +2 range)
- Repetition disabled
- Frequency 0.1 to 0.15 (If the Model gets stuck on a topic.)
- Presence 0.05
Reasoning enabled.
If you have no access to some sampler settings, the provider will use their default. So don't worry about it too much. 😘
Updated
Pre-History / System Prompt
Post-History / Jailbreak
Legacy
Pre-History / System Prompt
Post-History / Jailbreak