AMALIA-9B-0626-DPO

AMALIA-9B-0626-DPO

Writes and answers in European Portuguese as an open 9B chat model from Portuguese universities, tuned with supervised fine-tuning and DPO. Made for developers and researchers who need a European Portuguese language model.

Made by amalia-llm on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: AMALIA is an open-source language model aimed at European Portuguese. This DPO version has 9.2B parameters and follows supervised fine-tuning with preference tuning. Portuguese universities and research centres built it together, with government funding, aiming to be effective, sovereign and transparent.

Who it suits: It suits developers, researchers and organizations that need a language model that handles European Portuguese well, and want to inspect how it was trained.

Industry | Who uses it | How they use it | Example
Software | App developers | Build chat or writing features in European Portuguese | Add a Portuguese assistant to a product
Public sector | Government teams | Use a locally controlled Portuguese model | Draft replies to citizen questions
Research | NLP researchers | Study a national-language model and its training data | Compare the SFT and DPO versions
Publishing | Writers and editors | Draft and revise European Portuguese text | Rework a paragraph in a more formal tone

If you work in other languages only, or you need up-to-the-minute news, you can skip it, since its knowledge has a cutoff.

Where to use it: It runs wherever you can load a Safetensors text-generation model, on your own computer or server.

• Python tools that load Hugging Face language models
• Windows, macOS or Linux machines for downloading the files
• Servers that host chat models for an app

Use case | Who | What they get back
European Portuguese chat | Developers | Conversational answers in Portuguese
Instruction following | Product teams | Replies shaped by clear instructions
Long-context work | Researchers | Support for inputs up to 32k tokens

How to set it up: You need Python and enough memory for a 9.2B model, which is about 18.3 GB in BF16. The card part shown here gives no run steps beyond the download.

1. Install the Hugging Face hub tools.
2. Download the model files.
3. Load them with a tool that runs Safetensors language models.

pip install -U huggingface_hub
hf download amalia-llm/AMALIA-9B-0626-DPO

How to use it: Prompt the model in European Portuguese, with clear instructions. The card shown here has no run command, so follow the model card on Hugging Face for the exact chat format.

Example: You ask the model, in European Portuguese, to explain a topic in simple words. It replies in Portuguese text. The card describes this chat use but shows no sample output.

For newcomers and experts: This is a chat model that was built for European Portuguese instead of general English. A first test is to ask a short question in Portuguese and read the answer.

Pre-training extended EuroLLM training to cover more European Portuguese, with a maximum sequence length of 32k tokens. SFT came first, then DPO on pairs that include on-policy responses from the SFT model. DPO aims to reduce hallucination, toxicity and drift from instructions.

Pre-train on European Portuguese data
Run supervised fine-tuning
Run preference tuning with DPO
Prompt the model in Portuguese

Good to know: Both post-training phases used 64 NVIDIA H100 GPUs, and the card gives their running time.

Title: Training time of each phase on 64 H100 GPUs
Supervised fine-tuning: 76 hours
Preference tuning with DPO: 12 hours

• Data: pre-training used Arquivo.pt web data, EuroLLM pre-training data, long-context samples from Stack-v2 and synthetic long-context data.
• Datasets: the SFT and DPO datasets are released.
• Knowledge cutoff: June 2024.
• License: Apache 2.0.
• Cost to run: none named.
• Data it reaches: the text you send, handled on your own machine.
• Not for: the part of the card shown lists no exclusions.

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 9.2B
• Weights: about 18.3 GB in BF16; about 4.6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,340
• Likes: 21
• License: apache-2.0
• Languages: pt
• Based on: amalia-llm/AMALIA-9B-0626-SFT
• Quantized versions: 12
• Tags: safetensors, llama, text-generation, conversational

Links

• Model: https://huggingface.co/amalia-llm/AMALIA-9B-0626-DPO
• Files: https://huggingface.co/amalia-llm/AMALIA-9B-0626-DPO/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download amalia-llm/AMALIA-9B-0626-DPO

Then follow the usage in its model card. Nothing in this kit runs by itself.