CPU-friendly q4_k_m quantization of the merged Qwen2.5-7B-Instruct + adelie-qwen-roleplay-v2-lora checkpoint. Ships in a single 4.4 GB file — runs on a laptop without a GPU.

Companion to the LoRA adapter at ramyun/adelie-qwen-roleplay-v2-lora. The LoRA repo is for FP16 + GPU mounting; this GGUF is for laptops / end-users without CUDA.

How to use

llama.cpp / llama-cpp-python

bash

pip install llama-cpp-python --only-binary=:all: \
    --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu

huggingface-cli download ramyun/adelie-qwen-roleplay-v2-gguf \
    qwen-roleplay-v2.q4_k_m.gguf --local-dir ./models

python

from llama_cpp import Llama

llm = Llama(
    model_path="./models/qwen-roleplay-v2.q4_k_m.gguf",
    n_ctx=4096,
    n_threads=8,
)
out = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": "당신은 판타지 세계의 냉소적인 잡화상 주인입니다."},
        {"role": "user", "content": "할인 좀 안 돼요?"},
    ],
    temperature=0.7,
)
print(out["choices"][0]["message"]["content"])

Or mount via the AdelieAI engine

bash

huggingface-cli download ramyun/adelie-qwen-roleplay-v2-gguf \
    qwen-roleplay-v2.q4_k_m.gguf \
    --local-dir models/ours/qwen-roleplay-v2-gguf

MODEL_PATH=models/ours/qwen-roleplay-v2-gguf/qwen-roleplay-v2.q4_k_m.gguf \
PYTHONUTF8=1 .venv/Scripts/uvicorn core.api.app:app --port 8770

Quantization details


Source	merged FP16 (`Qwen2.5-7B-Instruct + adelie-qwen-roleplay-v2-lora`)
Method	llama.cpp `convert_hf_to_gguf.py` → `quantize` to `q4_k_m`
Original size	14.5 GB (FP16)
Quantized size	4.4 GB (3.25× compression)
Runtime VRAM	0 (pure CPU)
Runtime RAM	~5 GB

Style preservation verified: 4/5 persona prompts + 5/5 general prompts pass under greedy decoding (see recipe.md).

Recommended sampling

temperature=0.7 and top_p=0.9. Greedy decoding (temperature=0) sometimes diverges on q4_k_m — known quantization artifact, mitigated by sampling.

License

Inherits the Tongyi Qianwen License v1 from the base model (Qwen/Qwen2.5-7B-Instruct).

Citation

bibtex

@software{adelieai_2026,
  title  = {AdelieAI: a persona engine for small-deployment LLMs},
  author = {ramyun},
  year   = {2026},
  url    = {https://github.com/southglory/AdelieAI}
}