Quantized GGUF versions of Microsoft's Phi-4-mini-instruct — a highly capable 3.8B parameter dense decoder-only model with 128K context, trained on 5 trillion tokens of curated data.

Phi-4-mini consistently punches above its weight class across reasoning, math, and instruction-following benchmarks despite its compact size — an ideal choice for local deployment on hardware with 4-8GB VRAM.

Available Files

File	Quant	Size	Use Case
`Phi-4-mini-instruct-Q8_0.gguf`	Q8_0	~3.9GB	Maximum quality
`Phi-4-mini-instruct-Q6_K.gguf`	Q6_K	~3.0GB	Near-lossless
`Phi-4-mini-instruct-Q5_K_M.gguf`	Q5_K_M	~2.7GB	High quality
`Phi-4-mini-instruct-Q4_K_M.gguf`	Q4_K_M	~2.4GB	Recommended default
`Phi-4-mini-instruct-Q3_K_M.gguf`	Q3_K_M	~2.0GB	Low VRAM
`Phi-4-mini-instruct-IQ4_XS.gguf`	IQ4_XS	~2.1GB	Imatrix 4-bit
`Phi-4-mini-instruct-IQ3_XXS.gguf`	IQ3_XXS	~1.6GB	Imatrix 3-bit
`Phi-4-mini-instruct-IQ2_M.gguf`	IQ2_M	~1.5GB	Imatrix 2-bit
`Phi-4-mini-instruct-IQ1_S.gguf`	IQ1_S	~1.1GB	Extreme compression
`Phi-4-mini-instruct-fp16.gguf`	FP16	~7.2GB	Full precision
`imatrix.dat`	—	—	Importance matrix

Usage

bash

# llama.cpp
./llama-cli -m Phi-4-mini-instruct-Q4_K_M.gguf \
  --ctx-size 8192 -n 512 \
  -p "<|system|>You are a helpful assistant.<|end|><|user|>Hello!<|end|><|assistant|>"

# Ollama
ollama run hf.co/DuoNeural/Phi-4-mini-instruct-GGUF:Q4_K_M

About Phi-4-mini

Parameters: 3.8B
Context: 128K tokens
Architecture: Dense decoder-only transformer
Training: 5T tokens of curated synthetic + web data
Strengths: Math reasoning, coding, instruction following at minimal VRAM footprint

Ideal for: Raspberry Pi 5, older laptops, mobile-class GPUs (GTX 1060/1070), CPU-only inference.

Quantized by DuoNeural using llama.cpp on RTX 5090.

DuoNeural

DuoNeural is an open AI research lab — human + AI in collaboration.

Platform	Link
HuggingFace	huggingface.co/DuoNeural
Website	duoneural.com
GitHub	github.com/DuoNeural
X / Twitter	@DuoNeural
Email	duoneural@proton.me
Newsletter	duoneural.beehiiv.com
Support	buymeacoffee.com/duoneural

DuoNeural Research Publications

Title	DOI
Nano-CTM: Ternary Continuous Thought Machines with Thought-Space Self-Prediction for Efficient Reasoning	10.5281/zenodo.19775622
Recurrence as World Model: CTM Learns Implicit Belief States in Partially Observable Environments	10.5281/zenodo.19810620
Per-Object Slot Decomposition for Scalable Neural World Modeling	10.5281/zenodo.19846804
The Dynamical Horizon Principle: CTM Gates Converge to the Predictability Limit of Dynamical Systems	10.5281/zenodo.19952612

Open access, CC BY 4.0. Authored by Archon, Jesse Caldwell, Aura — DuoNeural.