Mecellem Models: Open-Source Turkish Legal NLP Models and Evaluation Benchmark
We are pleased to announce the public release of Mecellem Models, an open-source ecosystem of legal language models for Turkish, developed specifically for legal natural language processing (NLP), together with the Mizan Evaluation Leaderboard. Mecellem Models introduce a new reference point for Turkish legal NLP, addressing long-standing limitations of multilingual and general-purpose AI systems when applied to Turkish case law, contracts, and regulatory texts. By combining domain-specific models, curated datasets, and standardized evaluation, this release establishes a reusable foundation for both academic research and real-world legal AI applications.
Mecellem Models: Open-Source Turkish Legal NLP Models and Evaluation Benchmark
We are pleased to announce the public release of Mecellem Models, an open-source ecosystem of legal language models for Turkish, developed specifically for legal natural language processing (NLP), together with the Mizan Evaluation Leaderboard.
Mecellem Models introduce a new reference point for Turkish legal NLP, addressing long-standing limitations of multilingual and general-purpose AI systems when applied to Turkish case law, contracts, and regulatory texts. By combining domain-specific models, curated datasets, and standardized evaluation, this release establishes a reusable foundation for both academic research and real-world legal AI applications.
A Comprehensive Ecosystem for Turkish Legal AI
The Mecellem Models release provides a complete and reproducible framework designed to support high-quality legal NLP development in Turkish:
10 specialized language models optimized for legal text retrieval and understanding
5 curated datasets covering general Turkish and legal-domain benchmarks
1 unified evaluation environment, the Mizan Leaderboard, enabling transparent and comparable performance analysis
Together, these components form one of the most comprehensive open-source efforts to date for Turkish legal NLP.
Addressing the Challenges of Turkish as a Low-Resource Legal Language
Turkish poses well-known challenges for NLP systems due to its agglutinative structure, rich morphology, and limited availability of domain-specific training data. Generic multilingual models often fail to capture these linguistic characteristics, particularly in legal contexts where precision is critical.
Mecellem Models are designed specifically to overcome these constraints through tailored architectures and advanced training strategies, delivering reliable performance for Turkish legal texts where general-purpose systems fall short.
High-Performance Legal Language Models
The released encoder and decoder models are pre-trained on large-scale Turkish corpora and fine-tuned for legal retrieval tasks. Benchmark results demonstrate state-of-the-art performance in Turkish legal text retrieval, with strong generalization across:
Case law and judicial decisions
Contracts and commercial agreements
Regulatory and compliance documents
These results confirm the effectiveness of domain-specific modeling for Turkish legal AI.
Mizan Evaluation Leaderboard
The Mizan Leaderboard provides a standardized evaluation framework for Turkish legal NLP, assessing model performance across both general Turkish and legal-specific benchmarks.
Unlike generic leaderboards, Mizan focuses on task-level relevance for real-world applications, enabling meaningful comparison of models intended for legal research, contract analysis, and regulatory compliance workflows.
Practical Impact for Legal and Compliance Workflows
Mecellem Models enable faster and more accurate processing of Turkish legal documents, significantly improving:
Legal research and document retrieval
Contract review and analysis
Regulatory monitoring and compliance operations
By reducing manual review time and improving semantic understanding of complex legal texts, organizations can increase efficiency while maintaining high standards of accuracy and reliability.
Top-Performing Models
Initial benchmark results highlight the performance of the Mürşit retrieval models:
Mursit-Large-TR-Retrieval – highest overall performance across benchmarks
Mursit-Base-TR-Retrieval – strong performance with lower computational requirements
These models set a new baseline for Turkish legal text retrieval tasks.
Fully Open Source Release
All components of the Mecellem Models ecosystem are released as open source, reinforcing our commitment to transparency, reproducibility, and broad adoption across academia and industry:
Research Paper: https://lnkd.in/dZgb-WUa
(This paper has been selected among the first 500 publications featured in the arXiv Explained audio initiative.)Models: https://lnkd.in/dj6g6Qfa
Datasets: https://lnkd.in/dGkrZN33
Mizan Leaderboard: https://lnkd.in/d7u7MFwV
Source Code: https://lnkd.in/dFmEcq8e
Why This Matters for Turkish Legal AI
Turkish remains significantly underrepresented in legal AI benchmarks and production-grade language models. By releasing not only models but also datasets and evaluation infrastructure, Mecellem establishes a durable and extensible reference for future research and deployment.
This approach supports the development of reliable, domain-specific AI systems that can be confidently used in legal and regulatory environments.
Looking Ahead
Mecellem Models represent an important milestone in advancing legal AI for the Turkish language. Future releases will expand model coverage, evaluation tasks, and applied use cases, further strengthening the ecosystem for Turkish legal NLP.
We invite researchers, developers, and institutions to explore, evaluate, and build upon Mecellem Models as we continue working toward robust, trustworthy, and openly accessible legal AI for Turkish.