Polyglot

Polyglot is an initiative to close the linguistic divide in NLP by developing efficient and accessible foundation models for low-resource languages.

While recent breakthroughs in generative AI have been driven by large-scale foundation models, these advances have largely benefited high-resource languages, leaving many underrepresented languages behind. The current deep learning paradigm—heavily reliant on massive datasets and computing power—has unintentionally widened this gap, making it harder for speakers of low-resource languages to access and shape AI technologies that reflect their linguistic and cultural identities.

Polyglot addresses this imbalance by creating tools, models, and datasets that support open, sustainable, and inclusive AI development. We aim to empower researchers and communities working with low-resource languages through high-quality open-source resources, enabling them to build and fine-tune language models tailored to their needs.

Recent Publications 📚

• Tucano 2 Cool: Better Open Source LLMs for Portuguese | GitHub | Collection | Paper |
• Raising Bars, Not Parameters: LilMoo Compact Language Model for Hindi | GitHub | Collection | Paper |
• LilTii: A 0.6B Bengali Language Model that Outperforms Qwen | GitHub | Collection | Blog |
• ViTucano: A Portuguese Vision Assistant | GitHub | Collection |
• Tucano: Advancing Neural Text Generation for Portuguese | GitHub | Collection | Paper |
• TeenyTinyLlama: open-source tiny language models trained in Brazilian Portuguese | GitHub | Collection | Paper |

News 🚀

• [05/03/2026] Tucano2, LilMoo, and LilTii are released! 🦜 🐮 🐯
• [24/07/2025] Peer-reviewed article "Tucano: Advancing Neural Text Generation for Portuguese" is published in Patterns, with all models and datasets released on Hugging Face.
• [13/01/2025] We release ViTucano, a pair of vision assistants natively pretrained in Portuguese (ViTucano-1b5-v1, ViTucano-2b8-v1).
• [13/01/2025] We release the datasets used to pretrain and fine-tune the ViTucano models: ViTucano-Pretrain and ViTucano-SFT.
• [29/11/2024] Tucano is mentioned on Deutsche Welle: "Cientistas criam maior banco de dados em português para IA".
• [27/11/2024] Tucano video presentation at the C4AI (USP) [available on YouTube].
• [12/11/2024] "Tucano: Advancing Neural Text Generation for Portuguese" is published as a preprint on ArXiv, with all models and datasets released on Hugging Face.

Community Contributions 🤝

• Demo on how to run inference on ViTucano.
• Demo on how to run inference on Tucano.
• Demo on how to create a simple Chat UI for Tucano using Gradio.
• Tucano OpenVINO is a ported version of Tucano-2b4-Instruct optimized for Intel openVINO inference technology.

Acknowledgments

Polyglot is a project funded by the Federal Ministry of Education and Research (BMBF) and the Ministry of Culture and Science of the State of North Rhine-Westphalia (MWK) as part of TRA Sustainable Futures (University of Bonn) and the Excellence Strategy of the federal and state governments.

We also gratefully acknowledge access to the Marvin and Bender clusters, hosted by the University of Bonn, and maintained by the university's High Performance Computing Team. We also appreciate the support team that maintains the Bonn Analysis Facility (BAF) for its constant support and maintenance of the infrastructure we all share.

In addition to these local computing resources, we gratefully acknowledge the Gauss Centre for Supercomputing e.V. (www.gauss-centre.eu) for funding this project by providing computing time on the GCS Supercomputer JUPITER Booster at the Jülich Supercomputing Centre.