ansulev/Darwin-28B-REASON

Darwin-28B-REASON: Enhanced Scientific Reasoning Model

Darwin-28B-REASON is a 27.6 billion parameter standalone model, part of the Darwin family by FINAL-Bench / Darwin Research Team, specifically engineered for advanced reasoning tasks. It is built upon the Darwin-28B-Opus base model and integrates two core techniques:

Key Capabilities & Techniques

• Reasoning-Trace Distillation (RTD): This process distills complete reasoning chains from a mathematical corpus into the model, enhancing its ability to handle long-form, multi-step scientific reasoning while maintaining bilingual support (English, Korean, Chinese, Japanese).
• Darwin-DELPHI Test-Time Engine: A proprietary inference-time engine that performs multi-sample cross-validation, re-examination of uncertain responses, and iterative self-critique to converge on a consensus answer. This engine is not stored in the model weights but is crucial for its peak performance.
• Exceptional Reasoning Performance: Achieves 89.39% on the GPQA Diamond benchmark (198 PhD-level science questions) when combined with Darwin-DELPHI, making it a top-tier model for graduate-level scientific reasoning.
• Large Context Window: Supports a substantial context length of 262,144 tokens, facilitating long-chain reasoning.

Recommended Use Cases

• Graduate-level STEM reasoning and science qualifying exams.
• Mathematical problem-solving (e.g., MATH, AIME-style problems).
• Complex multi-step chain-of-thought tasks.
• Code generation and debugging.
• Bilingual reasoning with strong support for English and Korean, and secondary support for Chinese and Japanese.