Salamandra 7B

Salamandra is a suite of open-source decoder-only large language models developed by the Language Technologies Unit of the Barcelona Supercomputing Center, released in 2, 7, and 40 billion parameter sizes. The models were trained from scratch on a highly multilingual corpus covering 35 European languages plus code, curated exclusively from open-access data — a deliberate stance for open science. Instruction-tuned variants trained on public-domain instruction data are released alongside the base models for chat applications.

Limitations: as models trained only on openly licensed data, they trail frontier commercial models on English-centric benchmarks and complex instruction following; the 7B model is the practical sweet spot, with the 40B variant demanding serious hardware. Fine-tuning code and evaluation scripts are public, which is the real differentiator for researchers.