Another critical step has been taken in the field of artificial intelligence in Turkey. The open-source decision-support model Mizan-27B, trained specifically for the Turkish legal system, has been made available to users. The project, implemented by developer Alican Kiraz, stands out as a large language model (LLM) that has specialized in its field by being trained with tens of thousands of legal documents, from the Constitution to precedents. Aiming to keep information within national borders for both individual and corporate purposes, Mizan-27B also appeals to a wide user base with its low system requirements.

• Turkish Law-Specific Data Set: Trial Balance-27B; It was trained with over 21,000 documents consisting of the Constitution, legislation, precedents, and academic studies.

• High Performance: Built on the Qwen3.6-27B platform, the model shows a significant performance increase compared to the base model in benchmark tests.

• Wide Hardware Support: Accessible in 4-bit MLX and GGUF formats compatible with Apple Silicon (M series) chips and Nvidia and AMD graphics cards.

The Era of Artificial Intelligence in the Legal Sector

Artificial intelligence technologies are moving beyond general use and entering a phase of sectoral specialization. Especially in sensitive, highly terminological, and legislation-based fields such as law, general-purpose language models can sometimes be insufficient or produce inaccurate information. Developed with the aim of providing a solution to this need, Mizan-27B was designed to fully adapt to Türkiye’s legal dynamics.

The model, which underwent approximately 100 hours of training, serves as a decision-support mechanism in legal processes. It aims to significantly speed up time-consuming processes such as resource searching, text analysis, and precedent review for lawyers, academics, law students, and members of the judiciary.

Emphasis on Data Security and Domestic Infrastructure

According to information provided by developer Alican Kiraz, a hybrid approach with RAG (Retrieval-Augmented Generation) architecture was adopted in the project’s data-generation pipeline. The dataset, created by analyzing more than 21,000 documents, ensures that the model provides accurate and balanced legal outputs.

Another noteworthy aspect of the project is its data security and localization dimension. Instead of sending sensitive legal data to servers abroad, running it on local devices or in-house servers offers a strategic advantage in terms of data security. Sharing the model as open source reduces external dependence while also contributing to keeping information within national borders.

How to Use Mizan-27B and What are the System Requirements?

Mizan-27B is shared in optimized formats for easy access by developers and end-users. The model supports 4-bit quantized MLX and GGUF document formats. Thanks to this:

• It runs with high efficiency on Mac computers with Apple Silicon (M1, M2, M3, M4 etc.) processors thanks to the MLX format.

• Users with Nvidia and AMD GPUs can easily run the model on their standard computers or local servers via the GGUF format.

Mizan-27B, which produces much more successful results in Turkish legal texts and analyzes compared to the Qwen3.6-27B model it is based on in performance tests, is also a valuable milestone in terms of the development of the open source artificial intelligence ecosystem in Turkey.