EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Edge-Ready Romanian Language Models: Training, Quantization, and Deployment

T. A. Diac, P. F. de Viana, A. F. Neagoe, A. Oprea, M. C. Raportaru, A. Nicolin-Żaczek

AI · 06.02.2026 · 7 (2) · 61

DOI: 10.3390/ai7020061 ↗

Kurz gefasst

Fachartikel in der Zeitschrift AI, online erschienen am 06.02.2026. Verfasst von T. A. Diac und 5 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 38 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Romanian Ministry of Education and Research under Romanian National Core Program LAPLAS VII, European Union.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

We present RoBaseLM-S (125 M) and RoBaseLM-M (260 M), two compact Romanian decoder-only language models trained from scratch on a 4.3 B-token curated corpus. Architecturally, they follow a modern LLaMA-style recipe with pre-norm RMSNorm, rotary position embeddings, SwiGLU feed-forward blocks, grouped-query attention, and 4 k-token context windows. We release both full-precision (FP16) and post-training 5-bit (Q5_K_M) checkpoints in GGUF format for lightweight local inference. The 5-bit variants fit under 500 MB and generate text in real time on a Jetson Nano 4 GB, enabling fully offline Romanian text generation on consumer-grade edge hardware. We evaluate the models intrinsically (multi-domain perplexity across news, literary prose, poetry, and heterogeneous web text) and extrinsically (LaRoSeDa sentiment classification and RO-STS sentence similarity). Relative to Romanian GPT-2–style baselines at similar parameter scales, RoBaseLM-S and RoBaseLM-M reduce perplexity substantially, e.g., from 30.7 to 15.9 on our held-out news split. The 5-bit post-training quantized checkpoints remain within FP16 performance across all reported tasks. To our knowledge, these are the first Romanian small language models explicitly optimized for long-context inference, post-training quantization, and low-power on-device deployment.

Förderung

  • Romanian Ministry of Education and Research under Romanian National Core Program LAPLAS VII
  • European Union

Zitieren

APA

Diac, T. A., Viana, P. F. D., Neagoe, A. F., Oprea, A., Raportaru, M. C., & Nicolin-Żaczek, A. (2026). Edge-Ready Romanian Language Models: Training, Quantization, and Deployment. AI, 7(2), 61. https://doi.org/10.3390/ai7020061

Vancouver

Diac TA, Viana PFD, Neagoe AF, Oprea A, Raportaru MC, Nicolin-Żaczek A. Edge-Ready Romanian Language Models: Training, Quantization, and Deployment. AI. 2026;7(2):61. doi:10.3390/ai7020061

BibTeX
@article{Diac2026,
  author = {Diac, T. A. and Viana, P. F. de and Neagoe, A. F. and Oprea, A. and Raportaru, M. C. and Nicolin-Żaczek, A.},
  title = {Edge-Ready Romanian Language Models: Training, Quantization, and Deployment},
  journal = {AI},
  year = {2026},
  volume = {7},
  number = {2},
  pages = {61},
  doi = {10.3390/ai7020061}
}
Kostenlos starten →