From Data to Device: ELMOD An Efficient German-First 2.7B Language Model for Mobile Inference

2026-07-27

AI Summary

This work addresses the challenges of deploying large language models for German on resource-constrained mobile devices, where the language’s complex morphology, compound words, and orthographic rules pose significant hurdles. To this end, the authors propose ELMOD, a 2.7B-parameter German language model specifically optimized for on-device inference. By integrating tailored data preprocessing, education-quality–oriented filtering, and an automatic rewriting mechanism—combined with a compact Transformer architecture and efficient inference optimizations—ELMOD achieves state-of-the-art performance among sub-3B-parameter models on German-language tasks, rivaling that of 7B-scale models. The entire training process requires only 55k H100 GPU hours, demonstrating a highly efficient use of computational resources while maintaining strong linguistic competence.

Abstract

We present ELMOD - Efficient Language Model for On-Device Deployment - a compact (2.7B) German language model designed for efficient inference on resource-constrained hardware. ELMOD was trained on a limited computational budget (55k H100 GPU hours) using exclusively publicly available data. We developed a suite of German-specific data pre-processing, which differ from English-oriented counterparts in their handling of morphological variation, compounding, and orthographic conventions. Furthermore, we introduced a quality filtering and rephrasing step, which increased the instructional quality of the data, improved performance during the annealing phase, and reduced overall compute requirements. Thanks to our architectural model and data choices, including prefiltering, our educational-quality filtering and rephrasal to raise the educational-quality, ELMOD is the strongest performer in its size class (<3B), matching the performance of 7B-parameter models in German.

Problem

• on-device inference
• German language model
• resource-constrained hardware
• efficient deployment
• limited computational budget

Innovation

• on-device inference
• German-specific preprocessing
• data quality filtering
• efficient language model
• morphological compounding