FINAL-Bench/Darwin-28B-KR
Darwin-28B-KR is a 28 billion parameter multimodal language model developed by VIDRAFT, specialized for Korean language tasks. Built on the Qwen3.5 architecture, it supports Korean expression, understanding, and reasoning, alongside strong English reasoning and multimodal capabilities (image/video understanding). This model serves as a foundational base for future domain-specific Korean models in fields like law, medicine, and finance, offering a 32768 token context length.
Loading preview...
Darwin-28B-KR: Korean-Specialized Multimodal LLM
Darwin-28B-KR is a 28 billion parameter multimodal language model developed by VIDRAFT, designed with a strong focus on the Korean language. It represents the second generation of the Darwin family, inheriting capabilities from Qwen3.5-27B and undergoing an evolutionary merge process from Darwin-27B-Opus and Darwin-27B-KR.
Key Capabilities
- Superior Korean Language Proficiency: Excels in Korean understanding, generation, and complex reasoning, including CSAT/PSAT-level tasks.
- Robust English Reasoning: Maintains strong English reasoning abilities, inherited from the Darwin-28B-Opus lineage.
- Multimodal Understanding: Supports image and video comprehension, preserving capabilities from its parent models.
- English-Korean Code-Switching: Demonstrates strong performance in mixed-language contexts.
- Foundational Model: Serves as the base for upcoming domain-specific Korean models (e.g., Legal, Medical, Finance, Code).
Good For
- General Korean Conversation and Q&A: Ideal for natural and accurate interactions in Korean.
- Korean Cultural and Factual Knowledge: Provides reliable responses on Korean history, culture, and legal topics.
- Complex Korean Reasoning: Suitable for tasks requiring advanced logical inference in Korean.
- Multimodal Applications: Capable of analyzing images/videos and generating descriptions or insights in Korean.
- English-Korean Translation and English Reasoning: Effective for cross-lingual tasks and pure English reasoning.
This model achieves a CLIcK score of 0.786 on the K-AI leaderboard, demonstrating competitive performance among Korean-specialized models. It requires approximately 55 GB of VRAM for bfloat16 inference, making NVIDIA Hopper (H100/H200) or Blackwell (B200) GPUs recommended.