News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgSalamandra

Salamandra

Open multilingual language models for European languages

Salamandra is a family of open-source decoder-only language models developed by the Barcelona Supercomputing Center. Released in February 2025, it comes in three sizes and supports 35 European languages. The project aims to promote open science by sharing weights, training scripts, and data curation details under a permissive licence.

Share on WhatsApp

  • Trained on 12.875 trillion tokens of open data
  • Supports 35 European languages and 92 coding languages
  • Released under permissive Apache 2.0 licence
  • Available in 2B, 7B, and 40B parameter sizes
  • Trained on MareNostrum 5 supercomputer in Barcelona

What it is

Salamandra is a suite of large language models created by the Barcelona Supercomputing Center. It includes three variants with 2 billion, 7 billion, and 40 billion parameters. Each size has a base version for general generation and an instruction-tuned variant for chat applications. The models are designed to handle text in 35 European languages and code. They are built as decoder-only transformers and trained from scratch on a curated corpus of open-access data.

How it came about

The technical report for Salamandra was submitted to arXiv in February 2025. The development team, led by researchers at the Barcelona Supercomputing Center, aimed to demonstrate strong multilingual capabilities. The project emphasizes open science by releasing not just the model weights but also the training scripts and configuration files. This approach allows other researchers to reproduce the results and build upon the work. The models were trained on the MareNostrum 5 supercomputer.

Inside the model

The architecture uses a decoder-only transformer design with SwiGLU activation functions and RMS normalization. It employs grouped query attention to improve efficiency. Training took place on the MareNostrum 5 EuroHPC supercomputer using NVIDIA’s NeMo Framework. The dataset comprises 12.875 trillion tokens from sources like Colossal OSCAR and FineWeb-Edu. The team adjusted data proportions to give more weight to Spain’s co-official languages, such as Catalan and Basque, while downsampling English and code.

What it can do

Evaluations show that Salamandra performs competitively against other open-source models of similar size. It handles tasks in 35 European languages and supports 92 programming languages. The instruction-tuned versions are optimized for conversation using the ChatML template. Preliminary experiments indicate potential for multimodal applications, though the primary focus remains on text. The models have not been aligned with reinforcement learning from human feedback to filter sensitive topics.

How to use it

All models are released under the Apache 2.0 licence, permitting both research and commercial use. The weights, training scripts, and evaluation code are publicly available on GitHub and Hugging Face. Users can access the base models for fine-tuning or the instruction-tuned checkpoints for direct chat applications. The team encourages responsible usage and compliance with local laws.

Safety and testing

The developers note that the instruction-tuned models have not undergone alignment via reinforcement learning to avoid sensitive subjects. Consequently, they may generate harmful or inappropriate content. The team describes the release as a proof-of-concept for instruction-following capabilities. They state that work is ongoing to enhance performance through further alignment techniques. Users are advised to conduct risk assessments before deploying the models in production.

Why it matters

Salamandra contributes to the open-source ecosystem by providing high-quality models for European languages. By sharing its full training pipeline and data curation strategy, the project supports transparency in artificial intelligence research. It offers an alternative to models dominated by English data, promoting linguistic diversity in large language model development.

Questions people ask about Salamandra

What languages does Salamandra support?

Salamandra supports text in 35 European languages, including Spanish, Catalan, Basque, and Galician. It also handles code in 92 different programming languages. The training data was specifically balanced to give higher representation to Spain’s co-official languages.

Is Salamandra free to use for commercial projects?

Yes, the entire Salamandra family is released under the Apache 2.0 licence. This permissive licence allows for both research and commercial use. Users must comply with current laws and regulations when deploying the models in any application.

How was Salamandra trained?

The models were trained from scratch on the MareNostrum 5 supercomputer using NVIDIA’s NeMo Framework. The training corpus consists of 12.875 trillion tokens from open-access sources like Colossal OSCAR and FineWeb-Edu, with adjusted proportions for specific languages.

Can Salamandra understand images?

The primary models are text-based. However, the technical report mentions preliminary experiments on multimodality as a proof-of-concept. These experiments showcase potential applications, but the main release focuses on text generation and instruction following.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Salamandra yet, so the facts on this page come from the sources listed above: huggingface.co/BSC-LT/salamandra-7b-instruct, arxiv.org/abs/2502.08489. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.