File	Quant	Size	Recommended use
`aksarallm-1.5b-native.f16.gguf`	F16	4.08 GB	lossless from safetensors
`aksarallm-1.5b-native.q8_0.gguf`	Q8_0	2.17 GB	near-lossless, ~2× smaller
`aksarallm-1.5b-native.q6_k.gguf`	Q6_K	1.77 GB	high quality, ~2.5× smaller
`aksarallm-1.5b-native.q5_k_m.gguf`	Q5_K_M	1.53 GB	good quality, ~3× smaller
`aksarallm-1.5b-native.q4_k_m.gguf`	Q4_K_M	1.35 GB	recommended default, ~4× smaller

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

Quant	Prompt eval (32 tok)	Generation (16 tok)
`q4_k_m`	17.2 tok/s	9.7 tok/s

So a 2.04B 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/aksarallm-1.5b-native-GGUF aksarallm-1.5b-native.q4_k_m.gguf --local-dir .
./llama-cli -m aksarallm-1.5b-native.q4_k_m.gguf -p "Indonesia adalah" -n 64

Quick start — Ollama

bash

huggingface-cli download AksaraLLM/aksarallm-1.5b-native-GGUF aksarallm-1.5b-native.q4_k_m.gguf Modelfile --local-dir .
ollama create aksara-aksarallm-1.5b-native -f Modelfile
ollama run aksara-aksarallm-1.5b-native "Lanjutkan: Indonesia adalah negara"

Source model

See AksaraLLM/aksarallm-1.5b-native 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 llama
All files listed above are reproducible from the source HF safetensors

Note on the from-scratch model

This is a llama-3-style decoder built and trained from scratch by the AksaraLLM project. It does not use the Qwen2 ChatML template — it expects a plain ### Instruksi: ... ### Jawaban: ... style prompt (set up automatically by the included Modelfile).