DeepSeek V3.2 Presets for SillyTavern
I tried V3.2 again last month after bouncing off it earlier in the year, and I am not going to lie, it caught me off guard. “Apparently all this model needed was a quality preset. My goodness can it cook” is the vibe in the megathreads right now, and I’m inclined to agree. The community has it ranked #3 overall (per the Freaky Frankenstein tier list as of June 2026), the FF creator uses it interchangeably with GLM 5.1 depending on scene type, and the prose is vivid in a way the front-page OpenRouter rankings don’t really capture. The one real limitation is the 163K context window, which is lower than the 1M-token models, so very long sessions can fall off. That’s a tradeoff, not a dealbreaker.
I haven’t seen enough long-session V3.2 reports yet to call this settled, but early feedback is very positive. Here’s what the community is reaching for:
Freaky Frankenstein MAX+ / BOLT+ – best fit for V3.2
The FF creator uses V3.2 as one of his primary testing models, and it shows. No special DeepSeek toggle required for V3.2 (that’s only needed for V4). The model’s zero-censorship stance and strong prose combine with FF’s anti-slop logic and NPC initiative prompting to produce some of the most praised output in the community. MAX+ for quality-first sessions, BOLT+ when you want speed without a quality drop. Either way, you’re getting the FF engine running on one of the models it was tuned against. GLM 5.2 Presets
Stab’s EDH
The tiered directive hierarchy works just as well on DeepSeek’s architecture as it does on GLM’s. For V3.2, the recommendation is to disable the Visual Toolkit, Web Development, Relationship Tracker, and Chaotic Thoughts color dialog toggles. Keeps the token footprint lean and focuses the model on narrative output. The five-tier structure (Tier 0 = your commands override everything) ensures your character rules, behavioral rules, and prose guidelines are all enforced in order before every response. If you like the structure Stab’s brings to GLM and want the same on DeepSeek, give it a run.
momoura’s peepsqueak / deepdeep-dropseek – DeepSeek specific
This one’s interesting. deepdeep-dropseek is an experimental variant that attempts to get DeepSeek R1 to think from within the character’s perspective rather than instructing from the outside. That’s a fundamentally different approach to most system-level presets, and it works on V3.2 as a baseline too. Recommended model per the README: DeepSeek R1 0528. The instructions are embedded directly in the README at the top of the preset inside SillyTavern, so you don’t have to dig. DeepSeek V4 Flash Presets
If you prefer a lightweight, writing-focused preset for DeepSeek that prioritizes first-person character immersion over heavy scaffolding, this is worth your time.
If you bounce off V3.2, change the preset before you change the model. The model is fine.
More Presets here.