Skip to content
Bora Haber
Markets
  • US dollar50.44up 0.05%
  • Euro55.34down 0.18%
  • British pound66.55down 0.11%
  • BIST 10011,477down 3.08%
  • Gold (gram)6,041down 3.55%
  • Gold (ounce)US$3,725down 3.6%
  • BitcoinUS$62,686down 1.61%
BREAKING

Alican Kiraz Open-Sources Mizan-27B-Decision, a Decision Model for Turkish Law

Developer Alican Kiraz released the weights of Mizan-27B-Decision, a model that classifies legal cases and returns probabilities instead of generating text, under a non-commercial license.

Nehir Gökçe

· 3 min read

Bronze Lady Justice statue with scales on a white desk, beside case files, a laptop and pens; a suited man sits behind.
Illustrative image: a Lady Justice statue and legal documents on a law office desk. Temsili görsel · Fotoğraf: https://kaboompics.com/ / Pexels

AI developer Alican Kiraz has released as open source the weights of Mizan-27B-Decision, a decision model he built for Turkish law. The model is on Hugging Face as "Mizan-27B-Turkish-Legal-Decision-Model". It is not a text-generating chat model; it is a classification model that distributes probabilities across options.

Kiraz built it on top of Mizan-27B, a Turkish-law language model he announced on August 1. He applied additional training (LoRA) and added a small decision layer called a "pointer head" on top.

Six questions in a single pass

When a user enters a case summary, the model gives probabilities on six points: the area of law (14 options), the main statute to apply (13 options), the judicial route (7 options), the outcome of the claim (justified, partly justified, not justified), whether a time limit or statute-of-limitations issue affects the result, and the complexity of the case (simple, medium, complex). A confidence value is returned with each answer. Users can define their own questions and options, though quality may drop for options not seen in training.

Test results

According to the model card, Mizan-27B-Decision reached 94.9 percent accuracy on a held-out test set of 1,000 records and 5,505 questions. However, this figure measures agreement with labels produced by another AI model, DeepSeek v4.1 flash, not by human lawyers. The card also states that no human quality-control sample is included.

Results are lower on 20 hard cases containing boundary, exception and time-limit traps. In this test, which Kiraz says he prepared with Opus 5.5 and Astra, the model fully matched the expected answer on 93 of 119 questions, or 78.2 percent. Counting acceptable alternative answers, the rate rises to 88.2 percent. The model card says errors cluster in a "general rule correct, exception/time limit missed" pattern and recommends human review, especially for claim outcome and time-limit questions.

Kiraz stresses that the model is not designed to replace a lawyer. The developer said:

Model hukuki tavsiye vermez; uzman incelemesini önceliklendirmek için tasarladım.

Alican Kiraz

Demo space and hardware

A Space has been opened on Hugging Face to try the model without installation. Users can pick one of the 20 ready-made hard cases or type their own case text and press the "Karar ver" (Decide) button. Because the Space runs on ZeroGPU, the first request can take 30 to 60 seconds. Users logged in with a Hugging Face account get a daily GPU quota of 5 minutes, PRO subscribers get 40 minutes, and anonymous use is limited to 2 minutes.

For local use, the bf16 weights need about 52 GB of GPU memory. On an NVIDIA H100, a full six-question request is answered in 397 milliseconds. The weights are offered under CC BY-NC 4.0, meaning they can be downloaded for free but not used commercially. The training data is not shared with the repository.

Background

According to Kiraz, the base of MizanLLM is Qwen3.8-27B from Alibaba's Qwen3.8 family. The model card and model tree on Hugging Face, however, show Qwen3.6-27B at the start of the chain. Mizan-27B, released in August, used the same base.

Kiraz's cybersecurity model Titus ranked first among open-source models on the CS-Eval leaderboard in June.

Science guides

Take a quiz, vote in a poll

Video news

Comments

Sign in to comment.

More on this topic

See all