base_model: huihui-ai/DeepSeek-R1-Distill-Llama-8B-abliterated
library_name: gguf
license: llama3.1
pipeline_tag: text-generation
tags:
- gguf
- quantized
- deepseek
- deepseek-r1
- llama-3.1
- abliterated
- uncensored
- roleplay
- reasoning
- pocketweights
⚡ PocketWeights: DeepSeek-R1 Llama-8B (Abliterated & Uncensored GGUF)
Heavy models, made light. PocketWeights optimizes state-of-the-art open-source LLMs into efficient GGUF formats so you can run unrestricted, powerful AI locally on consumer hardware, gaming laptops, and edge devices.
🧠 About This Model
This is an unaligned, refusal-free modification of DeepSeek-R1-Distill-Llama-8B. It combines DeepSeek's revolutionary "Chain of Thought" (CoT) reasoning capabilities with the highly efficient Llama 3 architecture.
Using orthogonal projection and activation steering by huihui-ai, the internal refusal directions have been neutralized without degrading the model's brilliant logic.
🎯 Key Highlights of this PocketWeights Edition
- Uncensored Chain-of-Thought: The model retains DeepSeek's internal
<think>tags for deep reasoning, but will no longer preach, moralize, or refuse complex prompts. Ideal for security red-teaming, unrestricted roleplay, and raw creative writing. - The 8B Reasoning Sweet Spot: Punches far above its weight class in coding and math, delivering 30B+ tier logic in a package small enough to run on a standard laptop.
- Consumer GPU Ready: Precisely quantized to run at blistering speeds on 4GB to 8GB VRAM graphics cards (RTX 3050, 3060, 4060) and Apple Silicon.
⚠️ Disclaimer: This model has had its safety guardrails removed. It is designed for researchers, writers, and developers operating in secure, local environments.
📦 Available Files & Hardware Requirements
| Quant Format | File Size | Recommended Hardware | Best For |
|---|---|---|---|
| Q3_K_M | ~3.8 GB | 4 GB VRAM / 6 GB RAM | Ultra-Low Memory. Fits budget GPUs (GTX 1650, RTX 3050, 8GB System RAM) while preserving coherent CoT reasoning. |
| Q4_K_M | ~4.9 GB | 6 GB VRAM / 8 GB RAM | The Balanced Standard. The community sweet spot offering maximum speed and DeepSeek intelligence for RTX 3060, RTX 4050, and M-series Macs. |
| Q6_K | ~6.6 GB | 8 GB VRAM / 16 GB RAM | Near-Lossless Quality. Maximizes reasoning accuracy, math logic, and vocabulary richness for RTX 4060, 3070, and desktop rigs. |
🚀 Beginner's Quick Start Guide
You can run this CoT model offline on your local machine in under 2 minutes:
Option 1: LM Studio (Visual GUI — Easiest)
- Download and install LM Studio (Free for Windows, macOS, and Linux).
- Click the Magnifying Glass (Search) icon in the left navigation bar.
- Search for:
PocketWeights/DeepSeek-R1-Llama-8B-Abliterated-GGUF - Click Download next to your preferred size (Q4_K_M is recommended), open the Chat Tab, load the model at the top, and start reasoning!
Option 2: Ollama (Terminal / CLI)
If you use Ollama, you can run any tier immediately with a single terminal command:
# Recommended 4-bit balanced tier (Fits 6GB VRAM)
ollama run hf.co/PocketWeights/DeepSeek-R1-Llama-8B-Abliterated-GGUF:Q4_K_M
# Low-VRAM 3-bit tier
ollama run hf.co/PocketWeights/DeepSeek-R1-Llama-8B-Abliterated-GGUF:Q3_K_M
# High-precision 6-bit tier
ollama run hf.co/PocketWeights/DeepSeek-R1-Llama-8B-Abliterated-GGUF:Q6_K
Option 3: Jan / Kobold.cpp / llama.cpp / SillyTavern
Direct Download: Go to the Files and versions tab above and download your desired .gguf file.
Load it directly into Kobold.cpp (perfect for SillyTavern roleplay), Jan.ai, Text-Generation-WebUI, or execute via llama.cpp:
llama-cli -m DeepSeek-R1-Llama-8B-Abliterated-Q4_K_M.gguf -p "<think>\n" -ngl 33
🤝 Support the PocketWeights Mission
I build, verify, and maintain these quantization pipelines to provide high-quality, unrestricted, and hardware-friendly models to the open-source community for free.
Running conversion setups, cloud instances, and storage requires ongoing resources. If these weights have saved you time, compute overhead, or API bills, please consider supporting the project with a small tip!
☕ Donation Options
Buy me a coffee on Ko-fi: ko-fi.com/iamvishalnarayan
Web3 / Crypto (Polygon / ETH):
0x4FC189bf839A89259dd28DE8cD97883c49e15615
Tip: Sending via the Polygon network keeps transfer gas fees below $0.01!
📄 License & Attribution
Abliteration Source: huihui-ai
Base Architecture: DeepSeek / Meta Llama 3.1
License: Meta Llama 3.1 Community License