metadata

base_model: microsoft/phi-4
license: mit
license_link: https://huggingface.co/microsoft/phi-4/resolve/main/LICENSE
language:
  - en
pipeline_tag: text-generation
library_name: transformers
model_creator: Microsoft
model_name: phi-4
quantized_by: Second State Inc.

Phi-4-GGUF

Original Model

microsoft/phi-4

Run with LlamaEdge

LlamaEdge version: v0.16.0 and above

Prompt template

Prompt type: phi-4-chat

Prompt string

<|im_start|>system<|im_sep|>
{system_message}<|im_end|>
<|im_start|>user<|im_sep|>
{user_message}<|im_end|>
<|im_start|>assistant<|im_sep|>

Context size: 16000

Run as LlamaEdge service

wasmedge --dir .:. --nn-preload default:GGML:AUTO:phi-4-Q5_K_M.gguf \
  llama-api-server.wasm \
  --prompt-template phi-4-chat \
  --ctx-size 16000 \
  --model-name phi-3-mini

Run as LlamaEdge command app

wasmedge --dir .:. --nn-preload default:GGML:AUTO:phi-4-Q5_K_M.gguf \
  llama-chat.wasm \
  --prompt-template phi-4-chat \
  --ctx-size 16000

Quantized GGUF Models

Name	Quant method	Bits	Size	Use case
phi-4-Q2_K.gguf	Q2_K	2	5.55 GB	smallest, significant quality loss - not recommended for most purposes
phi-4-Q3_K_L.gguf	Q3_K_L	3	7.93 GB	small, substantial quality loss
phi-4-Q3_K_M.gguf	Q3_K_M	3	7.36 GB	very small, high quality loss
phi-4-Q3_K_S.gguf	Q3_K_S	3	6.50 GB	very small, high quality loss
phi-4-Q4_0.gguf	Q4_0	4	8.38 GB	legacy; small, very high quality loss - prefer using Q3_K_M
phi-4-Q4_K_M.gguf	Q4_K_M	4	9.05 GB	medium, balanced quality - recommended
phi-4-Q4_K_S.gguf	Q4_K_S	4	8.44 GB	small, greater quality loss
phi-4-Q5_0.gguf	Q5_0	5	10.2 GB	legacy; medium, balanced quality - prefer using Q4_K_M
phi-4-Q5_K_M.gguf	Q5_K_M	5	10.6 GB	large, very low quality loss - recommended
phi-4-Q5_K_S.gguf	Q5_K_S	5	10.2 GB	large, low quality loss - recommended
phi-4-Q6_K.gguf	Q6_K	6	12.0 GB	very large, extremely low quality loss
phi-4-Q8_0.gguf	Q8_0	8	15.6 GB	very large, extremely low quality loss - not recommended
phi-4-f16.gguf	f16	16	29.3 GB

Quantized with llama.cpp b4450.