license: apache-2.0
base_model: DavidAU/gemma-4-31B-it-Grand-Horror-X-INTENSE-HERETIC-UNCENSORED-Thinking
pretty_name: Sensix Paite 31B Master 4bit
language:
- pck
- en
pipeline_tag: text-generation
tags: - gemma-4
- thinking
- uncensored
- paite
- regional-language
- 4bit
- unsloth
Sensix Paite 31B Master (4-bit Quantized)
This repository provides the 4-bit quantized version of the Sensix Paite 31B Master model. This model is a high-performance derivative of the Gemma-4-31B architecture, specifically optimized for the Paite language through Continued Pre-Training (CPT). It integrates advanced reasoning, a dedicated thinking channel, and an uncensored response framework into a memory-efficient 4-bit format.
Model Description
The Sensix Paite 31B Master is designed to deliver high-IQ outputs and linguistic fluency in Paite without the high VRAM requirements of the full 16-bit model. It was created by merging a high-rank (r=128) LoRA adapter into the base weights and subsequently quantizing the result to 4-bit precision.
Key Modifications
Architectural Fix: Prior to training, the original Gemma 4 clippable linear layers were unwrapped and converted to standard linear modules. This ensures that the weights are fully compatible with quantization and inference engines without loss of structural integrity.
Linguistic CPT: The model has been trained on a comprehensive Paite corpus, including formal syntactical data and modern vocabulary updates (2025-12-13), allowing for high-fidelity translation and instruction following in the Paite language.
Thinking Framework: The model maintains its native internal reasoning channel. It can "think" through complex queries before providing a final output, ensuring logical consistency and depth.
Uncensored Heretic Persona: The model inherits an unfiltered, direct-response style. It is designed to follow instructions literally and avoids the standard refusal patterns found in safety-aligned models.
Technical Specifications
- Base Model: Gemma 4 31B (Instruction Tuned / Grand Horror)
- Quantization: 4-bit (Optimized for VRAM efficiency)
- Parameters: 31 Billion
- Context Window: 4096 Tokens
- Inference Method: Thinking / Reasoning-enabled chat
- Primary Languages: Paite (pck) and English (en)
Implementation and Usage
This 4-bit version is optimized for use with the Unsloth library or compatible 4-bit Transformers loaders.
Loading the Model
from unsloth import FastLanguageModel
import torch
model_id = "sensix-zo/Gemma-4-31B-Paite-Master-4bit"
model, tokenizer = FastLanguageModel.from_pretrained(
model_name = model_id,
max_seq_length = 4096,
dtype = torch.bfloat16,
load_in_4bit = True,
device_map = "auto",
)
FastLanguageModel.for_inference(model)
Prompting Protocol
To utilize the internal thinking process, ensure the chat template is applied with thinking enabled:
messages = [
{"role": "user", "content": "Paite pau in, education thupi dan gelh in."}
]
prompt = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=True
)
inputs = tokenizer(text=prompt, return_tensors="pt").to("cuda")
with torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=1000,
temperature=0.8,
top_p=0.95
)
print(tokenizer.decode(outputs[0], skip_special_tokens=False))
Performance and Memory
The 4-bit quantization significantly reduces the VRAM footprint, making it possible to run this 31B parameter model on consumer-grade hardware or mid-tier enterprise GPUs (e.g., RTX 3090/4090 or A6000).
- Estimated VRAM for Inference: ~18GB - 22GB
- Quantization Method: 4-bit BitsAndBytes / Unsloth
Disclaimer and Ethics
This model is UNCENSORED. It is designed to be a highly capable reasoning engine that does not refuse user instructions based on standard safety guardrails. Users are solely responsible for the content generated by the model. By using this model, you agree to comply with your local laws and regulations regarding AI-generated content. The developers are not liable for any misuse or harmful content produced by the model.