Deepseek-V4-Flash-SillyTavern-Preset

Deepseek V4 Flash abliterated by huihui is actually an insane model for roleplay.

• It can run locally if you give your inheritance to Corsair
• It eagerly engages with any sort of roleplay
• It is decently creative and proactive
• I find the writing style less sloppy than other models

But it has big problems:

• A short attention span
• Except during the first answer, it thinks for a short time (about three little paragraphs in chinese), even with reasoning_effort: max
• The thinking almost always impersonate the user, which is not helpful
• It doesn't follow instructions well, and doesn't follow character chards
• It forgets, it doesn't plan in advance, etc...

I made a chat template preset for SillyTavern that forces the model to think properly (recall, trying to impersonate characters, plan the next turn, follow rules), and it works much better than the original model trying to figure out what it's supposed to do. I really recommend trying this model, and this preset will reveal its true potential. Feel free to tweak the preset to your needs.

I recommend you add the following block to "connection profile > additional parameters > include body parameters"

chat_template_kwargs: {skip_special_tokens: false, reasoning_effort: max}
skip_special_tokens: false
min_p: 0.05
temperature: 0.70