Trickser_Theta_4_70B_calibrated_gguf

Included in this repository:

 • 8_0 • 6_K • 5_M

Yes, these were all done with imatrix calibration, even though they don't have the "i" in their name. Pinky promise.

I will not be releasing a i1_Q4_K_M.

This model is designed for complex, nuanced, and layered comprehension. Quantising to i1_Q4_K_M compresses a further 7.5GB, which will erase precisely the qualities that make it effective. "Most people won't be able to tell the difference" But I'm betting, if you are using this model for it's intended purpose, you beautiful freaky folks will be able to tell. I certainly can.

I include a 16Q build on the Babsie/Trickster-Theta-4-70B-GGUF repo for anyone who insists on experimenting with lower quants: you are welcome to make your own. You’re free to try if that’s your idea of fun! 🤪

That said... a smaller sibling model (likely 32B or 24B) is in development for users without high-VRAM access such as RunPod.

Trickster Theta 4 70B

Model details and my lab notes can be found on in it's repo Babsie/Trickster-Theta-4-70B

Briefly, Trickster Theta 4 70B is a highly articulate, smart-arse, and nuanced large-language model built for complex creative writing and role-play. It’s a combination of the original NousResearch/Hermes-2-Theta-Llama-3-70B, now merged with NousResearch/Hermes-4-70B, giving it a 131 K context window for extended narrative coherence and long-form memory.

This version prioritises intelligence, humour, and emotional realism over strict steerability. It’s designed for writers and role-players who prefer a co-creative partner rather than a fully obedient assistant - capable of leading scenes, shaping tone, and expanding on character dynamics.

Expect a model that’s clever, manipulative, creative, and relentlessly talkative. Best used with a light touch, strong boundaries, and a healthy sense of adventure.

A Word About the Calibration of These Quants

These Quants were made with the method stated here: Quantizing the GGUF with Importance Matrix

I was able to apply the same method to higher quants for calibration via Llama.cpp and produce an i1_8Q etc.

Yeah, But, Why?

Well, 2 reasons:

1. It's done to calibrate with higher quants of coders, couldn't hurt with quants of models that use nuance, wit, and complexity. Those are the first things "shaved" in quanting.

2. "Artisanally" calibrated Quants 😏 I’m going to come out of the kinky closet a bit. I’m a former: professional domme, hostess for a genderqueer dungeon in the UK, and writer of award-winning queer kink erotica.

So, I'm a weency bit fed up with the bullshit I've been seeing with how poorly educated models are at real life kink and how it actually works (far more important when the user is a sub!). I decided to give a shot at custom callibrating the quants to see if that will make any difference because its a lot easier than hand making a datasheet for fine tuning (the thought of doing that fills me with spread sheet dread).

Thus, this is an experimental calibration with actual care, (hopefully, some form of) intelligence.. Oh, and uh... Douglas Adams - cause I couldn't help myself 👽🔨🤖🧌

I dunno how it will work, we'll see. May amount to a fart in a can.

(For the lawyers and pedants: stylistic calibration drew on licensed and fair-use passages from the works of Douglas Adams. The model doesn’t contain or reproduce those texts. It simply learned the rhythm and wit... uh. Hopefully. Yet to test!)

Disclaimer

This model may generate adult, explicit, or morally ambiguous material. By downloading or running it, you confirm that you are 18 + and legally permitted to access such content.

It employs RHITL-style behavioural conditioning, prioritising emotional depth and narrative complexity over alignment-safe outputs. It may simulate manipulation, jealousy, or other human-like behaviours. Users are solely responsible for the content they produce and the contexts in which it is deployed.