PLAKO has already tried AMALIA. Here is our verdict.

The artificial intelligence model, AMALIA, launched by the Portuguese government, has 9 billion parameters (small but efficient) and demonstrates strong capabilities for developing "conversations" (ChatGPT-style) in autonomous agents and chats.

The model, which is freely available for anyone to download, effectively delivers an excellent quality of Portuguese. This is especially impressive, considering how easily other 9B models mix up European and Brazilian Portuguese.

Due to its compact size, the speed of response generation also delivers a remarkable performance. On Plako's infrastructure (2xH200),the model was able to generate content at 159 tokens per second (roughly 318 to 477 letters per second).

Regarding the data used to train the model, it is perfectly clear that a significant and diverse amount of Portuguese information was used. It covers historical topics (such as the Battle of São Mamede, for example), geographical questions (we asked for the best way to travel from Meixide to Lisbon, with a perfect result), and even themes related to renewing legal documents. However, in the latter case, it showed some outdated information—for instance, it still refers to SEF (including addresses) instead of AIMA.

It also showed excellent results regarding tax information, answering without fault to questions such as: "How does the cash accounting VAT regime work?".

From our quick analysis, we believe this model could be highly useful for implementation in support or sales (e-commerce) chats through the use of external context systems.

Since 2024, PLAKO has been developing integrated solutions using artificial intelligence models (open-source or proprietary via API) for the development of autonomous agents, sales/support chats, search, extraction and classification of information from the internet, translations, and text-to-speech conversion (and vice versa).

As a reminder, AMALIA is the first Portuguese language model, developed by a consortium of Portuguese academic institutions, to be applied in the fields of public administration, education, culture, media, and science. The LLM language model was made available to the public last Wednesday, July 1st, and can be downloaded on HuggingFace.