sovasoft/zora-v1.13
Zora v1.13 is an 8 billion parameter open language model developed by Sovasoft, built upon Qwen3-8B. It is specifically designed for 12 languages of the Balkans and Southeast Europe, focusing on honest, multi-perspective responses and admitting when it lacks information. This version emphasizes critical analysis, advanced logic, and graded evaluation, making it suitable for tasks requiring nuanced understanding in these regional languages.
Loading preview...
Zora v1.13: An Honest LLM for Southeast Europe
Zora v1.13, developed by Sovasoft, is an 8 billion parameter language model based on Qwen3-8B, specifically engineered for 12 languages of the Balkans and Southeast Europe. Unlike models that hallucinate, Zora is designed to be honest, providing multi-perspective views on contested topics and explicitly stating when it doesn't know an answer.
Key Capabilities & Differentiators
- Multilingual Focus: Optimized for Serbian, Croatian, Bosnian, Macedonian, Slovenian, Albanian, Montenegrin, Bulgarian, Greek, Turkish, Romanian, and Hungarian.
- Enhanced Reasoning & Analysis: Version 1.13 significantly improves critical analysis (12/12 on BalkanBench), advanced multi-step logic (7/12), and graded evaluation (3/12).
- Honest Refusal: Trained to provide structured, in-language "I don't know" responses rather than inventing facts, a core strength since v1.11.
- RAG Integration: Features Retrieval-Augmented Generation (RAG) for accessing current, authoritative data from a local knowledge base (Wikidata, Wikipedia, News Archive, EU Law, Statistics) with source transparency.
Use Cases & Considerations
- Good for:
- Applications requiring critical analysis and bias detection in Balkan languages.
- Tasks needing multi-step reasoning and structured responses.
- Educational tools for essay assessment and rubrics.
- Information retrieval where honesty and source transparency are paramount.
- Limitations:
- DETAIL Regression: v1.13 trades detailed answers for improved analysis, resulting in less elaborated responses (2/12 on BalkanBench).
- Factual Recall: Due to its 8B capacity, factual recall remains a structural weakness.
- Smaller Languages: Macedonian and Slovenian may exhibit lower consistency due to less training data.
Zora v1.13 represents a deliberate trade-off, prioritizing advanced reasoning and analytical capabilities over factual detail, making it a unique tool for specific applications in its target languages.