Lara Mini tops 36 models with up to 2B active params on 26 benchmarks (huggingface.co)

The newly released Lara Mini base model by Translated has emerged as a leader among 36 evaluated open base models, boasting a remarkable performance across various benchmarks despite having only 0.6 billion active parameters. It achieved an impressive average score of 57.0 over English, multilingual, math, and code tasks, outperforming models that were trained on significantly larger datasets. Its unique design allows it to process long contexts of up to 128k tokens while maintaining efficiency, utilizing a Gated DeltaNet and sliding-window attention mechanism that conserves computational resources. This model is particularly significant for the AI/ML community as it is both multilingual and supports 44 languages, covering all official EU languages, which broadens the accessibility of advanced AI capabilities. Additionally, the lightweight nature of the model—capable of generating around 1,100 characters per second on a single A100 GPU—makes it a practical option for researchers and developers concerned with speed and efficiency. While it is positioned as a base model intended for text continuation rather than interactive uses, its release under a non-commercial license encourages innovation and experimentation in the field.