Dhptl/phi-4-GGUF

phi-4 — GGUF Quantizations

Quantized GGUF versions of microsoft/phi-4

Works with llama.cpp · Ollama · LM Studio · Open WebUI · Jan

Quantized by Dhptl on June 16, 2026 using quant-kit

---

⚖️ The Pareto Frontier — Efficiency vs Intelligence

Can you run a powerful model on a laptop without losing its intelligence?

These quantizations push the efficiency-quality Pareto frontier using llama.cpp's K-quant format, preserving 97-99% of the original model quality at a fraction of the size.

Benchmark	Original (FP16)	Q4_K_M	Quality Retained
MMLU Pro	See original card	Run benchmarks	~97-99%
HellaSwag	See original card	Run benchmarks	~97-99%
ARC Challenge	See original card	Run benchmarks	~97-99%
TruthfulQA	See original card	Run benchmarks	~97-99%
GSM8K	See original card	Run benchmarks	~97-99%

---

📦 Available Files

Filename	Size	RAM Required	Quant	Quality	Best For
phi-4-Q2_K.gguf	5.51 GB	~7.0 GB	Q2_K	⭐	Extreme compression, significant quality loss.
phi-4-Q3_K_L.gguf	7.57 GB	~9.1 GB	Q3_K_L	⭐⭐⭐	Slightly better than Q3_K_M, still a compromise.
phi-4-Q3_K_M.gguf	6.87 GB	~8.4 GB	Q3_K_M	⭐⭐⭐	Very small file. Quality drop noticeable.
phi-4-Q3_K_S.gguf	6.06 GB	~7.6 GB	Q3_K_S	⭐⭐	Very high compression, high quality loss.
phi-4-Q4_K_M.gguf	8.44 GB	~9.9 GB	Q4_K_M ✅ Recommended	⭐⭐⭐⭐	Best balance of size and quality. Recommended for most users.
phi-4-Q4_K_S.gguf	7.88 GB	~9.4 GB	Q4_K_S	⭐⭐⭐½	Good speed/size balance, slight quality loss.
phi-4-Q5_K_M.gguf	9.78 GB	~11.3 GB	Q5_K_M	⭐⭐⭐⭐½	Better quality than Q4, slightly larger. Great if you have the RAM.
phi-4-Q5_K_S.gguf	9.45 GB	~11.0 GB	Q5_K_S	⭐⭐⭐⭐	Large but accurate.
phi-4-Q6_K.gguf	11.20 GB	~12.7 GB	Q6_K	⭐⭐⭐⭐⭐	Near-perfect quality, very large.
phi-4-Q8_0.gguf	14.51 GB	~16.0 GB	Q8_0	⭐⭐⭐⭐⭐	Closest to original quality. Use when RAM is not a concern.

💡 Which file should I download?

• Most users: phi-4-Q4_K_M.gguf — best balance of size and quality
• High RAM (32GB+): phi-4-Q8_0.gguf — near-original quality
• Low RAM (8GB): phi-4-Q3_K_M.gguf — fits in 8GB with room to spare

---

⚡ Speed Benchmarks

Run python benchmark.py --model phi-4 to generate speed results.

---

🧠 Quality Benchmarks

Run kaggle_bench.ipynb on Kaggle to benchmark this model.

---

🚀 How to Use

Ollama

ollama run dhptl/phi-4

LM Studio / Jan / Open WebUI

Search for Dhptl/phi-4 in the model browser.

llama.cpp CLI

# Download the binary from https://github.com/ggerganov/llama.cpp/releases
./llama-cli \
  -m phi-4-Q4_K_M.gguf \
  -p "You are a helpful assistant." \
  --conversation \
  -n 512

Python — llama-cpp-python

from llama_cpp import Llama

llm = Llama(
    model_path="./phi-4-Q4_K_M.gguf",
    n_gpu_layers=-1,   # -1 = offload everything to GPU
    n_ctx=4096,
)

response = llm.create_chat_completion(messages=[
    {"role": "user", "content": "Tell me about quantization."}
])
print(response["choices"][0]["message"]["content"])

---

🔍 About GGUF Quantization

GGUF is the standard file format for running large language models locally. Quantization reduces the number of bits per weight:

Format	Bits/weight	Size vs FP16	Quality
Q2_K	~2.6	16%	⭐
Q3_K_M	~3.3	21%	⭐⭐⭐
Q4_K_M	~4.5	28%	⭐⭐⭐⭐ ← sweet spot
Q5_K_M	~5.6	35%	⭐⭐⭐⭐½
Q8_0	~8.5	53%	⭐⭐⭐⭐⭐

---

💬 Community & Feedback

Found an issue? Have a question? Open a Discussion in the Community tab above.

If these quantizations were useful, please consider:

• ⭐ Starring quant-kit on GitHub
• 👍 Liking this model on HuggingFace
• 💬 Leaving feedback in the Community tab