Phi4-Reasoner-Uncensored-GGUF
Model Summary
Phi4-Reasoner-Uncensored-GGUF is an uncensored GGUF conversion of Microsoft's reasoning-focused Phi-4 Mini Reasoning model, released for local inference, research, experimentation, and open-ended instruction following.
This version aims to preserve the strong reasoning, mathematics, coding, analytical thinking, and multi-step problem-solving capabilities of the original model while reducing alignment restrictions and response filtering commonly present in safety-tuned releases.
The model is intended for users who prefer maximum output freedom and direct responses during local deployment.
Key Features
- 🧠 Strong reasoning capabilities
- 📐 Advanced mathematical problem solving
- 💻 Coding and debugging assistance
- 📚 Long-context support
- 🔓 Reduced alignment restrictions
- ⚡ GGUF format for llama.cpp-compatible runtimes
- 🖥️ Suitable for local deployment
- 🏠 Works with LM Studio, Ollama, KoboldCPP, Jan, Open WebUI, and llama.cpp
Base Model
This model is derived from:
Phi-4-mini-reasoning is a 3.8B parameter transformer model specifically trained for reasoning-intensive tasks and mathematical problem solving. Microsoft reports strong performance across benchmarks including AIME, MATH-500, and GPQA. (Hugging Face)
Modifications
Phi4-Reasoner-Uncensored-GGUF introduces the following changes:
- Removal or reduction of refusal behavior where possible
- Reduced safety filtering
- Increased willingness to answer controversial, fictional, speculative, and unrestricted prompts
- Preservation of reasoning-focused behavior
- GGUF conversion for efficient local inference
- Quantized variants for resource-constrained hardware
No claims are made that all alignment mechanisms have been completely removed.
Intended Use
Recommended
- Research
- Education
- Coding assistance
- Mathematical reasoning
- Creative writing
- Story generation
- Roleplay
- Simulation
- Agent frameworks
- Local AI assistants
- Experimental AI research
Not Recommended
- Medical diagnosis
- Legal advice
- Financial advice
- High-risk autonomous systems
- Safety-critical environments
Users are responsible for validating all outputs.
Context Length
| Feature | Value |
|---|---|
| Parameters | 3.8B |
| Context Length | 128K |
| Architecture | Decoder-Only Transformer |
| Vocabulary | 200K+ Tokens |
| Format | GGUF |
Based on the original Phi-4-mini-reasoning architecture. (Hugging Face)
Prompt Format
Chat Template
<|system|>
You are a helpful reasoning assistant.
<|end|>
<|user|>
Explain how binary search works.
<|end|>
<|assistant|>
Recommended Settings
temperature: 0.6
top_p: 0.95
min_p: 0.05
repeat_penalty: 1.05
max_tokens: 4096
Example Use Cases
Mathematics
- Algebra
- Calculus
- Statistics
- Proof generation
- Olympiad-style reasoning
Coding
- Python
- JavaScript
- C++
- Rust
- SQL
- Debugging
- Code explanation
Reasoning
- Logic puzzles
- Multi-step planning
- Research assistance
- Agent workflows
Creative Tasks
- Worldbuilding
- Character creation
- Fiction writing
- Interactive storytelling
Hardware Requirements
Approximate recommendations:
| Quant | RAM Requirement |
|---|---|
| Q4_K_M | 6-8 GB |
Actual requirements vary by context size and backend.
Limitations
Like all language models, this model may:
- Hallucinate facts
- Generate incorrect reasoning
- Produce inaccurate citations
- Reflect biases present in training data
- Generate offensive or controversial content
- Produce unsafe outputs if prompted
Users should independently verify important information.
License
This repository inherits the license and usage requirements of the original Microsoft Phi-4-mini-reasoning release.
Please review the original license before commercial deployment:
Original Phi-4-mini-reasoning License and Model Card
Acknowledgements
Special thanks to:
Created by: WithinUsAI
Model: Phi4-Reasoner-Uncensored-GGUF
Type: Reasoning LLM / GGUF
Status: Community Release
Version: 1.0