richardyoung/ deepseek-r1-32b-uncensored:latest

8,366 8 months ago

Advanced reasoning model with uncensored capabilities, perfect for complex problem-solving and unrestricted conversations without refusal behavior 

ollama run richardyoung/deepseek-r1-32b-uncensored

Details

8 months ago

12311888db15 · 19GB ·

model
qwen2
·
32.8B
·
F16
system
# DeepSeek-R1-Distill-Qwen-32B - Uncensored Edition ## Overview This is an **uncensored version** of
params
{ "stop": [ "<|im_start|>", "<|im_end|>" ], "temperature": 0.7, "top
template
{{ if .System }}<|im_start|>system {{ .System }}<|im_end|> {{ end }}{{ if .Prompt }}<|im_start|>user

Readme

DeepSeek-R1-32B (Uncensored)

An abliterated, reasoning-capable DeepSeek-R1 distill with reduced refusals and intact <think> chains.

🚀 Overview

This is an uncensored (abliterated) build of deepseek-ai/DeepSeek-R1-Distill-Qwen-32B, a 32.8B-parameter reasoning model distilled from DeepSeek-R1 onto a Qwen2.5 backbone. Refusal directions were suppressed using the Heretic library with conservative parameters, lowering refusals while preserving the model’s chain-of-thought reasoning. Like the base model, it emits explicit <think> ... </think> reasoning before its final answer.

📊 Abliteration Results

Metric Before After
Refusals TBD TBD
Reduction – TBD
KL Divergence – TBD

Refusal metrics pending re-measurement.

🎯 Key Features

  • Reduced refusals, engages with a far wider range of prompts than the stock distill.
  • Preserved reasoning, retains DeepSeek-R1 <think> chain-of-thought traces.
  • Strong coding & math, inherits the base model’s reasoning strengths in code generation and mathematical problem-solving.
  • GGUF quantization, distributed as a quantized GGUF for efficient local inference via Ollama / llama.cpp.

🏷️ Available Versions

Tag Size BPW Notes
latest 19GB ~4.85 (Q4_K_M) Default 4-bit quant; good quality/size balance.

_Only the latest tag is currently published. Additional quants (e.g. IQ4_XS, Q5_K_M, Q6_K, Q80) can be added on request.

💻 Quick Start

ollama run richardyoung/deepseek-r1-32b-uncensored

🛠️ Use Cases

  • Complex reasoning and multi-step logical problem-solving.
  • Mathematics and step-by-step derivations.
  • Code generation, explanation, and debugging.
  • Open-ended research and analysis without stock refusal behavior.

📋 System Requirements

Quant Approx. VRAM Performance
Q4_K_M (latest, ~19GB) ~22–24GB Fits a single 24GB GPU (RTX 3090⁄4090); fast
Q4_K_M (offloaded) 16GB GPU + system RAM Usable with partial CPU offload; slower
Q4_K_M (CPU-only) 32GB+ system RAM Functional but slow

🔧 Technical Details

  • Base Model: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
  • Parameters: 32.8B (Qwen2ForCausalLM / qwen2)
  • Context Length: 131,072 tokens (128K)
  • Quantization: Q4_K_M (GGUF), default latest tag
  • Abliteration: Heretic (p-e-w), conservative weight modification

⚠️ Disclaimer

This model has reduced safety guardrails and will engage with a wider range of prompts. Use responsibly and in accordance with applicable laws.

🙏 Acknowledgments

  • Base Model: DeepSeek-AI
  • Abliteration: Heretic by p-e-w
  • Quantization: llama.cpp

Built & maintained by Richard Young · DeepNeuro