Portugal unveiled its first open-source artificial intelligence model ‘Amalia,’ joining a growing push across Europe for greater AI sovereignty and reduced reliance on foreign providers.
…
The move follows similar initiatives in other European countries, including France and Germany, where governments have backed own AI companies such as Mistral AI and Aleph Alpha to provide alternatives to models developed by U.S. firms including OpenAI, Google and Anthropic.
…
The Portuguese model, dubbed Amalia in honour of the late fado icon Amalia Rodrigues, was developed by a consortium of Portuguese universities and research institutions, with government backing and €5.5 million ($6.26 million) in EU recovery funds.
Portugal’s Prime Minister Luis Montenegro told the launch event, “Europe’s strategic autonomy is today, perhaps more than ever, tied to AI. This model will enable us to face the coming decades with greater sovereignty and less dependence.”
Montenegro said Amalia would help boost productivity across the public and private sectors, including banking, insurance, telecommunications and industry, while ensuring security, adding: “We will continue to invest heavily in this project.”
…
The Amalia large language model, along with its training dataset and source code, is released under an open-source license.
Initial applications include a virtual guide for Portugal’s museums, decision-support tools for the Portuguese Navy, an AI-powered teaching assistant for lesson planning, and a digital assistant to help the state deliver public services to citizens.
Amalia also leverages Portugal’s investment in high-performance computing, including access to the Deucalion and MareNostrum 5 supercomputers, giving it the computing power needed to train and run large AI models.
…
Unlike commercial AI assistants such as ChatGPT, however, Amalia is not being launched as a consumer-facing chatbot.
Instead, it will serve as an open technological platform that public institutions, companies and researchers can build upon to create their own AI applications.
…
“The Torment Nexus is now Open Source” is not something to be proud of
Nice, actually open source, rather than ‘open weights’, or straight up closed like 'Open’AI (yes, they released one open weight model).
If I understand correctly, the training data isn’t actually opensource
- Data from Arquivo.pt, obtained after processing collections of public web pages and free books;
- Pre-training data from the EuroLLM model;
- Long-context data samples from Stack-v2;
- Synthetic data to improve long-context retention, filtered from public datasets (needle-llama3-16x8k, needle_32k_finetuning_dataset).
Sounds like they scraped “public” websites for data?
Well, Boo. Thanks for the correction. Hopefully it’s at least reproducible. Any serious open source effort needs high quality training sets, otherwise it’s back to just relying on the kindness of corpos (or their desire to backstab each other).
Full training data available, including source and processed. But I can’t find the code. Presumably it doesn’t do anything new technologically. https://huggingface.co/amalia-llm
Apertus is open source.
C’mon Europe! Distill Anthropic and OpenAI like the Chinese models and then open source your model, weights and training data!
deleted by creator
