FINAL-Bench/Darwin-28B-KR-Legal

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 26, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

FINAL-Bench/Darwin-28B-KR-Legal is a 28 billion parameter multimodal language model developed by VIDRAFT, based on the Qwen3.5 architecture. This model is specifically fine-tuned for the Korean legal domain, enhancing its capabilities in legal reasoning, consultation, and interpretation. It excels at analyzing Korean court precedents, citing legal provisions, and solving Korean bar exam questions, making it ideal for legal research and assistance.

Loading preview...

Darwin-28B-KR-Legal: Specialized Korean Legal AI

Darwin-28B-KR-Legal is a 28 billion parameter multimodal language model developed by VIDRAFT, specifically tailored for the Korean legal domain. Built upon the Darwin-28B-KR base model, which itself is a variant of Qwen3.5-27B, this model has undergone extensive fine-tuning using Korean legal data, including court precedents, legal Q&A, and 14 years of Korean bar exam questions.

Key Capabilities

  • Korean Legal Knowledge: Demonstrates strong proficiency in Korean civil, criminal, public, commercial, labor, real estate, and intellectual property law.
  • Legal Reasoning: Excels at analyzing court precedents, citing legal provisions, and interpreting legal texts.
  • Bar Exam Performance: Highly capable in solving Korean bar exam problems.
  • Multimodal: Supports multimodal inputs (image/video) alongside its strong language capabilities.

Use Cases

  • Legal Research & Summarization: Efficiently analyzes and summarizes court precedents and legal provisions.
  • Case Analysis: Provides insights into factual relationships, applicable laws, and potential judgments.
  • Legal Document Assistance: Aids in drafting legal documents such as contracts, complaints, and answers.
  • Legal Education: Supports law students in their studies, particularly with bar exam preparation.

This model is designed for reference and assistance; it is not a substitute for professional legal advice. It requires significant hardware, with a minimum of approximately 55 GB VRAM for bfloat16 inference, ideally on NVIDIA Hopper or Blackwell GPUs.

Finetunes
1 models