From Data to Device: ELMOD An Efficient German-First 2.7B Language Model for Mobile Inference
ELMOD is a 2.7B German-first model built to run on constrained devices while competing with much larger German models.
The paper says it was trained with 55k H100 GPU hours using only public data. Its preprocessing is tailored to German morphology, compounds, and orthography rather than English-centric defaults. The authors also added quality filtering and rephrasing to improve instruction data and reduce compute needs. They claim ELMOD is the strongest sub-3B model in its class and matches 7B-parameter models in German. ArXiv · AI/CL/LG's note
The paper says it was trained with 55k H100 GPU hours using only public data. Its preprocessing is tailored to German morphology, compounds, and orthography rather than English-centric defaults. The authors also added quality filtering and rephrasing to improve instruction data and reduce compute needs. They claim ELMOD is the strongest sub-3B model in its class and matches 7B-parameter models in German. ArXiv · AI/CL/LG's note
score 5