CYFRAGOVPL/PLLuM-4B-base-2512

Hugging Face
VISIONConcurrency Cost:1Model Size:4.3BQuant:BF16Ctx Length:32kPublished:Jan 15, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

The PLLuM-4B-base-2512 is a 4.3 billion parameter large language model developed by the PLLuM consortium and continued by HIVE AI, specialized in Polish with additional English data. Based on Google's Gemma-3-4b-pt, it is designed to generate contextually coherent text and assist in tasks like question answering and summarization, particularly excelling in Polish public administration contexts. The model was trained on high-quality Polish and English corpora, refined through extensive instruction tuning, and aligned using a unique Polish preference corpus.

Loading preview...

PLLuM-4B-base-2512: Polish Language Model

The PLLuM-4B-base-2512 is a 4.3 billion parameter model from the PLLuM family, developed by the PLLuM consortium and later HIVE AI. It is built upon Google's Gemma-3-4b-pt and is specifically designed for the Polish language, incorporating additional English data for broader generalization. The model's development involved rigorous data collection, including large-scale, high-quality Polish and English text corpora, with a focus on cleaning and deduplication.

Key Capabilities

  • Specialized Polish Language Understanding: Optimized for generating contextually coherent text in Polish.
  • Extensive Instruction Tuning: Fine-tuned with approximately 70k manually curated Polish "organic instructions," 33k programmatically derived instructions, 15k RAG-style context-processing instructions, and 45k synthetic, context-aware instructions.
  • Preference Learning: Aligned using ~60k manually annotated Polish preference pairs to ensure safer, balanced, and contextually appropriate responses, even for sensitive topics.
  • Strong Performance: Achieves top scores on custom benchmarks relevant to Polish public administration and state-of-the-art results in broader Polish-language tasks.
  • RAG Optimization: Trained to perform well in Retrieval-Augmented Generation (RAG) settings, providing answers based on provided documents with citations.

Good for

  • General Language Tasks: Text generation, summarization, extraction, and question answering in Polish.
  • Domain-Specific Assistants: Particularly effective for applications within Polish public administration, legal, or bureaucratic domains.
  • Research & Development: Serves as a foundational model for downstream AI applications requiring strong Polish language capabilities.