OpenLLM-France released the multilingual model Luciole-23B-Base

The collective OpenLLM-France presented a new language model Luciole-23B-Base. The release is focused on working with texts in French and English. The model contains 23 billion parameters and was trained on a corpus of 5 trillion tokens. The developers provided access to the model's weights under the free Apache 2.0 license, which allows its use in commercial and research projects without restrictions. During internal tests, the model demonstrated high performance, receiving a rating of 6 stars according to the results of a five-hour evaluation cycle. The architecture is optimized to balance the quality of text generation and computational costs when deploying on standard server equipment. The release of Luciole-23B-Base expands the availability of specialized tools for the French-speaking segment of the AI market. The presence of an open license and a significant amount of training data make the model a tool for fine-tuning for specific natural language processing tasks.