---
license: apache-2.0
language:
- pt
base_model:
- amalia-llm/AMALIA-9B-0626-DPO
base_model_relation: quantized
pipeline_tag: text-generation
library_name: llama.cpp
tags:
- gguf
- llama.cpp
- quantized
- portuguese
- european-portuguese
- amalia
---

# AMALIA-9B-0626-DPO — GGUF

GGUF quantizations of [amalia-llm/AMALIA-9B-0626-DPO](https://huggingface.co/amalia-llm/AMALIA-9B-0626-DPO) for use with [llama.cpp](https://github.com/ggml-org/llama.cpp) and compatible runtimes (LM Studio, Ollama, Jan, etc.).

AMALIA is an open-source language model targeting European Portuguese, developed by a consortium of Portuguese universities and research centres and funded by the Government of Portugal. This is the DPO (preference-tuned) variant. See the [original model card](https://huggingface.co/amalia-llm/AMALIA-9B-0626-DPO) and the [technical report](https://arxiv.org/abs/2603.26511) for full details.

## Files

| File | Quant | Size | Notes |
| --- | --- | --- | --- |
| `AMALIA-9B-0626-DPO-Q4_K_S.gguf` | Q4_K_S | ~5.4 GB | 4-bit small, quantized directly from bf16 |
| `AMALIA-9B-0626-DPO-Q4_K_M.gguf` | Q4_K_M | ~5.6 GB | 4-bit medium, recommended balance of quality and size |
| `AMALIA-9B-0626-DPO-Q8_0.gguf` | Q8_0 | ~9.7 GB | 8-bit, near-lossless |
| `AMALIA-9B-0626-DPO-BF16.gguf` | BF16 | ~18.3 GB | full-precision original weights in GGUF |

Converted with `convert_hf_to_gguf.py` (llama.cpp b8470) from the original bf16 weights. BF16 and Q4_K_S come directly from the original weights; Q8_0 was converted directly from bf16 and Q4_K_M was requantized from that Q8_0 (Q8_0 is near-lossless relative to bf16). The chat template (ChatML with the AMALIA Portuguese system prompt) is embedded in the GGUF metadata.

## Usage

```bash
# llama.cpp CLI (chat template is picked up automatically)
llama-cli -hf layerx-labs/AMALIA-9B-0626-DPO-GGUF:Q4_K_M

# or with a local file
llama-cli -m AMALIA-9B-0626-DPO-Q4_K_M.gguf -cnv

# OpenAI-compatible server
llama-server -m AMALIA-9B-0626-DPO-Q4_K_M.gguf -c 8192
```

## License

Apache 2.0, same as the original model. All credit for the model goes to the [AMALIA project](https://amaliallm.pt/).