ELMOD 2.7B

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. This work was performed as research project in the scope of the compute time project ELMOD at NHR@FAU. 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 preprocessing, which differ from English-oriented counterparts in their handling of morphological variation, compounding, and orthographic conventions, to ensure high model performance, we developed a suite of German-specific data preprocessing. Furthermore, we introduced an 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 choices as well as our 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.

For more details, please refer to our paper!

Model Variants

• ELMOD 2.7B Base ← You are here
• ELMOD 2.7B Instruct

Model Performance

Comparison of German-capable base models with <= 3 parameters, averaged across German base tasks, among which ELMOD-2.7B performs best

Comparison of German-capable base models with 1B–8B parameters, averaged across German base tasks; models <=3 are hatched; ELMOD-2.7B performs on par with 7B models

Model Summary

ELMOD is a 2.7B dense transformer model.

Usage Example

from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig

tokenizer = AutoTokenizer.from_pretrained("fraunhofer-iis/elmod-2.7b-base")
model = AutoModelForCausalLM.from_pretrained("fraunhofer-iis/elmod-2.7b-base")
generation_config = GenerationConfig.from_pretrained("fraunhofer-iis/elmod-2.7b-base")
generation_config.max_length = 16
generation_config.pad_token_id = tokenizer.eos_token_id

inputs = tokenizer("Denn eben wo Begriffe fehlen,", return_tensors="pt")
inputs.pop("token_type_ids", None)
generated_ids = model.generate(**inputs, generation_config=generation_config)
print(tokenizer.decode(generated_ids[0], skip_special_tokens=False))

This models is also compatible with vLLM.

Citation

@misc{elmod-2.7b,
  title={From Data to Device: ELMOD. An Efficient German-First 2.7B Language Model for Mobile Inference},
  author={Darina Gold, Alexander Schwirjow, Viktor Haag, Viktor Hangya, Joel Schlotthauer, Fabian Küch und Luzian Hahn},
  year={2026},
  url={https://huggingface.co/fraunhofer-iis/elmod-2.7b-base}
}

Limitations

The generated content may not always be factually accurate, logically consistent, or free from biases present in the training data. These model should not be used as a definitive source of information. Any generated content should always be critically evaluated.

Legal Aspects

EU AI Act Transparency Documentation

Public Training Data Summary for GPAI-Models

Unlawful use (Third Party Rights)

Any use of ELMOD in a manner that infringes, misappropriates, or otherwise violates any third party’s rights, including but not limited to intentionally using ELMOD to generate outputs that infringe, misappropriate, dilute or otherwise violate copyrights is unlawful and prohibited by law.

Data Protection and Copyright Requests

We as Fraunhofer IIS maintain policies to ensure compliance with laws governing copyright material. If you believe your copyrighted material has been used without proper authorization, please contact: copyright-requests@iis.fraunhofer.de

To enable us to process your claim efficiently, please provide as much of the following information as possible in your correspondence:

We will act within reasonable time from receiving a complaint in a diligent and non-arbitrary manner. We reserve the right not to respond if (i) a request is manifestly unfounded or (ii) has already been addressed to an identical or similar complaint by the same rightsholder.

For removal requests of personally identifiable information (PII), please contact: datenschutz@iis.fraunhofer.de. Include the term "ELMOD" in the subject of your request.

License

This model is licensed under the [Creative Commons Attribution Non Commercial 4.0] (https://spdx.org/licenses/CC-BY-NC-4.0).

Acknowledgments

This work has been funded by the Free State of Bavaria in the DSgenAI project (Grant Nr.: RMF-SG20-3410-2-18-4). The authors gratefully acknowledge the scientific support and HPC resources provided by the Erlangen National High Performance Computing Center (NHR@FAU) of the Friedrich-Alexander-Universität ErlangenNürnberg (FAU) under the NHR project ELMOD: Efficient language models for on-device deployment (Grant Nr.: b239dc). NHR funding is provided by federal and Bavarian state authorities. NHR@FAU hardware is partially funded by the German Research Foundation (DFG) – 440719683.