File	Quant	Size	Recommended use
`Kiel-Pro-0.5B-v3.f16.gguf`	F16	0.99 GB	lossless from safetensors
`Kiel-Pro-0.5B-v3.q8_0.gguf`	Q8_0	0.53 GB	near-lossless, ~2× smaller
`Kiel-Pro-0.5B-v3.q6_k.gguf`	Q6_K	0.51 GB	high quality, ~2.5× smaller
`Kiel-Pro-0.5B-v3.q5_k_m.gguf`	Q5_K_M	0.42 GB	good quality, ~3× smaller
`Kiel-Pro-0.5B-v3.q4_k_m.gguf`	Q4_K_M	0.40 GB	recommended default, ~4× smaller

CPU benchmark (AMD EPYC 7763, 2 threads, AVX2)

Quant	Prompt eval (32 tok)	Generation (16 tok)
`q4_k_m`	36.7 tok/s	20.1 tok/s

So a 494M model at q4_k_m runs comfortably on a CPU laptop. Larger quants (q5_k_m, q6_k, q8_0) trade a bit of speed for better quality.

Quick start — llama.cpp

bash

huggingface-cli download AksaraLLM/Kiel-Pro-0.5B-v3-GGUF Kiel-Pro-0.5B-v3.q4_k_m.gguf --local-dir .
./llama-cli -m Kiel-Pro-0.5B-v3.q4_k_m.gguf -p "Indonesia adalah" -n 64

Quick start — Ollama

bash

huggingface-cli download AksaraLLM/Kiel-Pro-0.5B-v3-GGUF Kiel-Pro-0.5B-v3.q4_k_m.gguf Modelfile --local-dir .
ollama create aksara-kiel-pro-0.5b-v3 -f Modelfile
ollama run aksara-kiel-pro-0.5b-v3 "Apa ibukota Indonesia?"

Source model

See AksaraLLM/Kiel-Pro-0.5B-v3 for architecture, training data, eval results, and limitations.

Conversion provenance

Converted with convert_hf_to_gguf.py from llama.cpp
Quantized with llama-quantize from the same build
Architecture detected as qwen2
All files listed above are reproducible from the source HF safetensors