VIDraft/Darwin-28B-KOREA
VIDraft/Darwin-28B-KOREA is a 27 billion parameter dual-language model from VIDraft, optimized for both Korean and English reasoning. It is the first PERFECT parent pair merge in the Darwin series, combining the strengths of Darwin-28B-Opus for English reasoning and Rogue-28B-MIX for native Korean reasoning. This model achieves equivalent or superior performance across all evaluated areas compared to its parent models, making it suitable for complex bilingual inference tasks.
Loading preview...
Overview
VIDraft/Darwin-28B-KOREA is a 27 billion parameter model developed by VIDraft, representing the first "PERFECT parent pair" merge in their Darwin series. It combines two distinct parent models: FINAL-Bench/Darwin-28B-Opus, which excels in English reasoning and token efficiency, and ginigen-ai/Rogue-28B-MIX, known for its native Korean capabilities and deep Korean reasoning. The merge utilizes a per-layer linear interpolation algorithm, dynamically weighting the contributions of each parent model to achieve performance that is at least equivalent to, and often better than, either parent across all metrics.
Key Capabilities
- Bilingual Reasoning: Optimized for strong inference in both Korean and English.
- Performance Equivalence: Achieves 96.6% accuracy on 29 objective questions, matching its parent models, with no degradation in any area.
- Long-Chain Reasoning: Successfully inherits the long-chain reasoning patterns from the Korean-native parent model.
- Efficient Token Usage: Maintains a balanced average response token count (521) compared to its parents.
Good For
- Bilingual Applications: Ideal for use cases requiring robust reasoning and generation in both Korean and English.
- Complex Queries: Recommended for tasks where chain-of-thought reasoning is beneficial, especially with
max_tokensset to 1024 or higher. - Accurate Bilingual Output: Suitable when the primary goal is accurate answers, even if the reasoning process involves English thinking followed by Korean output.
Considerations
- For purely native Korean reasoning without any English thought patterns, the parent model Rogue-28B-MIX might be preferred.