MiguelOli/AMALIA-9B-0626-DPO-mlx-4bit

AMALIA-9B-0626-DPO — MLX 4-bit

MLX (Apple Silicon–native format) conversion of amalia-llm/AMALIA-9B-0626-DPO, the open LLM for European Portuguese (pt-PT) developed by NOVA, IST, IT and FCT (PROPOR 2026).

This is the DPO variant (instruct + preference alignment) — the most polished model in the family. Quantized to 4-bit to run fast and light on M1/M2/M3/M4 Macs.

Why this repo: at conversion time there was an MLX build only of the SFT variant (dgomes03) and a GGUF of the DPO variant (duarteocarmo), but no MLX build of the DPO variant. This repo fills that gap.

Usage

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("MiguelOli/AMALIA-9B-0626-DPO-mlx-4bit")

messages = [{"role": "user", "content": "Explica-me o que é o fado, em português de Portugal."}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)

response = generate(model, tokenizer, prompt=prompt, max_tokens=512, verbose=True)

Or from the command line:

mlx_lm.generate --model MiguelOli/AMALIA-9B-0626-DPO-mlx-4bit \
  --prompt "Qual a diferença entre 'está-se' e 'está a ser' em pt-PT?"

Conversion details

Base model	amalia-llm/AMALIA-9B-0626-DPO (9B, EuroLLM-9B continued pretraining + SFT + DPO)
Quantization	4-bit (4.5 bits/weight, default group size of mlx_lm.convert)
Tool	mlx-lm 0.31.3 (mlx_lm.convert --hf-path ... -q --q-bits 4)
Context	32K tokens
License	Apache-2.0 (inherited from the base model)

Tested on a Mac M4: ~18.75 tokens/sec generation, ~5.4 GB peak memory.

Usage tip

AMÁLIA's strength is authentic European Portuguese (phonetics, orthography, idioms, pt-PT vs pt-BR bias). Talk to it in pt-PT to get the most out of it. It is not the strongest model on raw factual knowledge — it is a 9B focused on the variant.

Credits

• Original model: the AMALIA project — NOVA, NOVA LINCS, Instituto de Telecomunicações, Instituto Superior Técnico, FCT. (amalia-llm)
• MLX conversion: this repo.