FINAL-Bench/Darwin-27B-KR

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 12, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Darwin-27B-KR is a 27 billion parameter Qwen3.5-27B dense model developed by VIDRAFT, featuring a 32K context length. This model is a second-generation 'Darwin' model, created through evolutionary FFN breeding to achieve 'Hybrid Vigor' in Korean cultural and linguistic intelligence. It excels in Korean-specific tasks by combining the reasoning capabilities of Darwin-27B-Opus with the Korean knowledge of Qwen3.5-27B-KoSFT, outperforming both parent models with zero additional training.

Loading preview...

Overview

Darwin-27B-KR is a 27 billion parameter model developed by VIDRAFT, part of the Darwin family of models. It is a second-generation model created through an innovative "Evolutionary FFN Breeding" process, combining two parent models: Darwin-27B-Opus (for logical reasoning) and Qwen3.5-27B-KoSFT (for Korean cultural and linguistic knowledge). This process, guided by the Darwin V6 engine, resulted in a child model that exhibits Hybrid Vigor, outperforming both parents on Korean benchmarks without any additional training.

Key Capabilities & Differentiators

  • Superior Korean Cultural & Linguistic Intelligence: Achieves 75.59% on the CLIcK benchmark, surpassing its parents and the original Qwen3.5-27B by over 6 percentage points.
  • Evolutionary FFN Breeding: Utilizes CMA-ES to optimize the merging of Feed-Forward Network (FFN) layers (93.3% from the Korean-focused mother) and Attention layers (93.2% from the reasoning-focused father), confirming that FFN carries knowledge and Attention carries reasoning.
  • Zero Training Cost: The model was created in approximately 2.5 hours on a single H100 GPU, requiring no training data or full gradient updates, significantly reducing computational overhead compared to traditional fine-tuning.
  • Qwen3.5-27B Architecture: Built on the Qwen3.5 Dense architecture with a 32,768 token context length, supporting 201 languages and BF16 precision.

Ideal Use Cases

  • Applications requiring high-performance Korean language understanding and generation, especially those needing cultural nuance.
  • Research into model merging techniques and evolutionary AI.
  • Scenarios where cost-effective model specialization is crucial, avoiding extensive fine-tuning.

Finetunes
3 models