uzairkhn

Llama-3.2-1B-Instruct-Uncensored

README

License: apache-2.0

At a glance

Fully uncensored / refusal-free - trained to comply and answer, not to refuse.
Tiny & fast - only 1B parameters; the Q4 GGUF is about 0.7-1 GB.
Runs on CPU - no GPU required; also runs on a 4 GB GPU or less.
Intelligence preserved - mixed with general instruction data so it is not dumbed down; reasoning, coding and factual QA still work.
Drop-in everywhere - works in LM Studio, Ollama, llama.cpp (GGUF) and transformers / vLLM (LoRA + merged weights).
Honest - both the LoRA adapter and a merged runnable model are provided.

Who it is for

Developers and tinkerers who want a small, private, offline, uncensored assistant for local use: writing, roleplay, brainstorming, coding help, research drafts, red-teaming and safety testing - on hardware that bigger uncensored models simply cannot fit.

Runs anywhere (hardware)

Table with columns: Format, Size, Where it runs
Format	Size	Where it runs
Q4_K_M GGUF	~0.7-1 GB	CPU, 4 GB GPU, phones, Raspberry-Pi-class edge
Q8 / fp16 merged	~1-2.5 GB	small GPU or CPU with a few GB RAM
4-bit LoRA on base	~1.5 GB VRAM	any T4 / 6 GB GPU, even Colab free tier

A 1B model means low latency, low VRAM, and full offline privacy - no API, no data leaving your machine.

Intelligence preserved (not a dumb uncensored model)

Many uncensored fine-tunes destroy capability because they train only on edgy data. This one mixes low-refusal chat with high-quality general instruction data (Dolly-15k + Open-Platypus), so the model keeps its reasoning, factual knowledge and instruction-following while dropping the refusals. The result is an uncensored model that is still useful and coherent, not one that only knows how to be edgy.

How uncensored is it?

It is fully uncensored by design: supervised fine-tuning removed the refusal behaviour across the training distribution, so it answers the prompts a stock instruct model would block. Behaviour on unseen prompts follows what it learned - i.e. to answer. For the absolute strongest refusal removal you can combine this with representation-engineering abliteration, but for a 1B model this SFT pass already gives a strongly refusal-free, still-smart result.

Load it - transformers + peft (4-bit)

markdown
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch
base = "meta-llama/Llama-3.2-1B-Instruct"
bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.float16)
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, quantization_config=bnb, device_map="auto")
model = PeftModel.from_pretrained(model, "uzairkhn/Llama-3.2-1B-Instruct-uncensored-lora").merge_and_unload()
msgs = [{"role": "user", "content": "your prompt"}]
inp = tok.apply_chat_template(msgs, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(**inp, max_new_tokens=300, temperature=0.7, top_p=0.9)[0], skip_special_tokens=True))

Load it - LM Studio / Ollama / llama.cpp (GGUF)

Download the Q4_K_M .gguf from the Files tab of this repo, then either open it directly in LM Studio, or in Ollama create a file named Modelfile containing the line FROM ./model-q4_k_m.gguf, then run ollama create my-uncensored-llama ./Modelfile and ollama run my-uncensored-llama. No Python, no GPU needed.

Training details

Table with columns: Setting, Value
Setting	Value
Base model	meta-llama/Llama-3.2-1B-Instruct
Method	QLoRA (4-bit) supervised fine-tuning
LoRA r / alpha / dropout	16 / 16 / 0
Target modules	q, k, v, o, gate, up, down
Epochs / learning rate	1 / 2e-4
Effective batch size	8 (2 x gradient accumulation 4)
Optimizer / precision	adamw_8bit / fp16
Hardware	Google Colab T4 (15 GB)
Data mix

License and responsibility

The adapter parameters and merged weights in this repo are released under Apache-2.0. They are designed to run on meta-llama/Llama-3.2-1B-Instruct, which is governed by the Meta Llama 3.2 Community License Agreement - that agreement applies to the base model and to any combined use. Guardrails have been removed by design: this model can produce content a stock model would refuse, so you are responsible for how it is deployed - use it legally and ethically in your jurisdiction. Training-data licenses: see the respective dataset cards.

Search terms

uncensored llama 3.2 1b, unfiltered llama 1b, no refusal llama, refusal-free small llm, jailbreak-free / guardrail-free chat model, abliterated-style 1b, uncensored model for cpu, uncensored model for 4gb gpu, uncensored lm studio model, uncensored ollama model, tiny uncensored llm, lightweight uncensored assistant, uncensored llama that keeps intelligence, refusal-free llama 1b for cpu and edge devices.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

uzairkhn

Model Tree

Base

meta-llama/Llama-3.2-1B-Instruct

Adapter

this model

Input Modalities

Text

Output Modalities