ricoz/meam-phi-4-mini-instruct-gguf

Phi-4-mini-instruct, Q4_K_M GGUF, for Meam.ai

This is Microsoft's Phi-4-mini-instruct, converted to GGUF and quantized to Q4_K_M with llama.cpp. Nothing else is changed: it isn't fine-tuned or merged with anything.

Meam.ai, a study tutor for Ontario students in Grades 7 to 12, downloads this file the first time it opens. The tutor runs on the device. The app keeps the file only if its SHA-256 matches the one below.

File	phi-4-mini-instruct-q4_k_m.gguf
SHA-256	0c544c02dfde5b595de204b6e2a9d80d4e2a47fbbc53904214aa9325b3546959
Size	2,493,840,704 bytes
Weights	microsoft/Phi-4-mini-instruct at revision cfbefacb99257ffa30c83adab238a50856ac3083
Converted with	llama.cpp at commit 171e8846b4af9766c354064cb776cb34a50f053f, the version llama.rn uses

Rebuild it yourself

The same steps give the same file, so you can check that it was made from Microsoft's weights:

git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
git checkout 171e8846b4af9766c354064cb776cb34a50f053f
cmake -B build -DGGML_METAL=OFF -DLLAMA_CURL=OFF -DCMAKE_BUILD_TYPE=Release
cmake --build build --target llama-quantize -j 8
python3 -m venv venv
venv/bin/pip install -r requirements/requirements-convert_hf_to_gguf.txt -e gguf-py
hf download microsoft/Phi-4-mini-instruct --revision cfbefacb99257ffa30c83adab238a50856ac3083 --local-dir hf \
  config.json generation_config.json model.safetensors.index.json model-00001-of-00002.safetensors \
  model-00002-of-00002.safetensors tokenizer.json tokenizer_config.json special_tokens_map.json \
  added_tokens.json vocab.json merges.txt README.md LICENSE
venv/bin/python convert_hf_to_gguf.py hf --outtype bf16 --outfile model-bf16.gguf
build/bin/llama-quantize model-bf16.gguf phi-4-mini-instruct-q4_k_m.gguf Q4_K_M
shasum -a 256 phi-4-mini-instruct-q4_k_m.gguf

Chat format

The file carries Phi-4-mini's own chat template:

<|system|>You are a helpful tutor.<|end|><|user|>How do I start?<|end|><|assistant|>

Generation ends at <|end|> or <|endoftext|>.

Licence

MIT, from Microsoft. LICENSE is Microsoft's licence file, copied unchanged. Phi-4-mini-instruct's own model card covers its training, intended uses and limits.