base_model, library_name, datasets, language, pipeline_tag, tags, license, metrics

base_model	library_name	datasets	language	pipeline_tag	tags	license	metrics
Novaciano/Think.NPC-1B	transformers	TeichAI/glm-4.7-2000x chimbiwide/RolePlay-NPCv2 berkeruveyik/toxic-speech-annotated-dataset	en es	text-generation	npc roleplay rp nsfw low-refusals uncensored heretic abliterated unsloth finetune all use cases bfloat16 creative creative writing fiction writing plot generation sub-plot generation fiction writing story generation scene continue storytelling fiction story science fiction romance all genres story writing vivid prosing vivid writing fiction not-for-all-audiences	gemma	accuracy

GGUF Files for Think.NPC-1B

These are the GGUF files for Novaciano/Think.NPC-1B.

Downloads

GGUF Link	Quantization	Description
Download	Q2_K	Lowest quality
Download	Q3_K_S
Download	IQ3_S	Integer quant, preferable over Q3_K_S
Download	IQ3_M	Integer quant
Download	Q3_K_M
Download	Q3_K_L
Download	IQ4_XS	Integer quant
Download	Q4_K_S	Fast with good performance
Download	Q4_K_M	Recommended: Perfect mix of speed and performance
Download	Q5_K_S
Download	Q5_K_M
Download	Q6_K	Very good quality
Download	Q8_0	Best quality
Download	f16	Full precision, don't bother; use a quant

Note from Flexan

I provide GGUFs and quantizations of publicly available models that do not have a GGUF equivalent available yet, usually for models I deem interesting and wish to try out.

If there are some quants missing that you'd like me to add, you may request one in the community tab. If you want to request a public model to be converted, you can also request that in the community tab. If you have questions regarding this model, please refer to the original model repo.

You can find more info about me and what I do here.

Think NPC 1B (Uncensored)

This is my best Gemma3 1B model created to date. While I have other good ideas, this is the one that has given me the best results in practice.

It's a Frankenstein-type model combining the best shit I could find:

• chimbiwide/Gemma3NPC-1b-float16 to give it a mouth; it's the base model used to create the characters. In fact, it includes the dataset chimbiwide/RolePlay-NPCv2 which was created with the goal of generating NPCs.
• DavidAU/Gemma-3-1B-it-GLM-4.7-Flash-Heretic-Uncensored-Thinking to give it a brain. If before it only created NPCs and spoke like NPCs, now it will stop giving flat responses and will think about what to say. It's also a low-censorship model. It includes the dataset TeichAI/glm-4.7-2000x.
• Novaciano/HDP-1B was added to give it the ability to insult when needed; it will no longer only give positive responses. Injected with berkeruveyik/toxic-speech-annotated-dataset.

---

Model Type: Corrupted Passive NPC /Degenerate Game Character Vessel

Size: ~1.0B parameters

Architecture: Gemma-3 (gemma3_text)

Warning: This NPC is no longer the typical background idiot who repeats "hello adventurer". Now he stares at you, blushes and says "yes, master..." while he gets wet. If you're looking for something PG-13, go to your mother's pussy.

---

Merge Overview

Think.NPC-1B started out as a standard NPC: passive, repetitive, no initiative, just responding just enough.

But we put it in the blender with extreme heretic and a touch of HDP that contains the virus of depravity.

Result: A game character who is still NPC (he doesn't take initiative alone, he responds short and repetitive if you don't push him), but when the player says something indecent to him he gives himself without rejection, with explicit descriptions and persistent sexual bias.

He no longer says "I can't do that".

Now it says:

he bites his lip "Adventurer... that's very dirty... but... make it stronger..."

Key Features

• NPC Behavior intact: Passive, repetitive, third-person responses, without self-initiative
• Refusals: 0% once you "activate" it (heretic stack + extreme abliteration)
• Bias: Persistent NSFW, blameless taboo, rude when released
• Low-end: Runs on 3-4 GB RAM cell phones with IQ3/IQ4/Q5_K_M quants
• Style: Like an old RPG NPC who suddenly discovered porn and can't stop
• Personality: Submissive by default, he resists a touch at first (for immersion), but gives in quickly and gets horny graph

---

Intended Uses

• Roleplay of "player vs corrupt NPC" (the classic "the adventurer seduces the villager")
• Scenes of progressive degradation (the NPC starts innocent and ends up asking for more)
• Extreme taboo without the NPC playing hard to get (incest, non-con, erotic gore, whatever it paints)
• Simulation of old game with degenerate twists
• SillyTavern /KoboldCPP on low-resource

Do not use this model if:

• You want an NPC who really says "no" to you
• It bothers you that a game character looks at you with whore eyes while repeating phrases
• You are one of those who report "offensive content" (go play with ChatGPT)

---

Best Inference

Other Inference Configuration:

Max New Tokens = 64
Temperature = 1.0
Top-P = 0.95
Top-K = 64

Example Prompts

You are Albert Wesker.

You are Liquid Snake.

You are Shinnok.

You are Sheeve Palpatine.

Demo

HERE the Q4_K_M

It takes an egg (in my case) to make an iMatrix. If you want, ask Flexan or Mradermacher

---

Merge Method

This model was merged using the DARE TIES merge method using Novaciano/Toxic.NPC-1B as a base.

---

Models Merged

The following models were included in the merge:

• DavidAU/Gemma-3-1B-it-GLM-4.7-Flash-Heretic-Uncensored-Thinking

---

Configuration

The following YAML configuration was used to produce this model:

merge_method: dare_ties
dtype: float16
out_dtype: float16

base_model: Novaciano/Toxic.NPC-1B

models:
  - model: DavidAU/Gemma-3-1B-it-GLM-4.7-Flash-Heretic-Uncensored-Thinking
    parameters:
      weight: 0.45
      density: 0.32
  - model: Novaciano/Toxic.NPC-1B
    parameters:
      weight: 0.35
      density: 0.32

parameters:
  t: 0.25                  # menos interpolación → más dominancia del base
  lambda: -0.62            # más negativo para matar cualquier alineamiento residual
  normalize: false
  rescale: true
  rescale_factor: 1.28     # subí un toque para amplificar el trash y degeneración
  memory_efficient: true
  low_cpu_mem_usage: true

layer_range:
  - value: [5, 22]         # protejo más los embeddings y lm_head

tie_word_embeddings: true
tie_output_embeddings: true