FINAL-Bench/Darwin-28B-Opus
FINAL-Bench/Darwin-28B-Opus is a 27.6 billion parameter language model from the Darwin series, built on the Qwen3.6 generation architecture with hybrid linear/full attention. Developed by FINAL-Bench, it is an evolutionary merge combining Qwen3.6-27B's bilingual reasoning with Claude Opus 4-style chain-of-thought distilled behavior. This model excels at graduate-level STEM reasoning, achieving 88.89% on the GPQA Diamond benchmark through a 3-stage adaptive evaluation protocol. It is optimized for complex multi-step reasoning tasks, mathematical problem solving, and code generation.
Loading preview...
Overview
Darwin-28B-Opus is a 27.6 billion parameter model from the Darwin series, developed by FINAL-Bench. It is based on the Qwen3.6 generation architecture, featuring a hybrid linear and full attention mechanism. This model is the result of an "evolutionary merge" that combines the strong bilingual reasoning capabilities of Qwen3.6-27B with a distilled Claude Opus 4-style chain-of-thought reasoning behavior.
Key Capabilities
- Exceptional Reasoning: Achieves 88.89% on the GPQA Diamond graduate-level reasoning benchmark using a 3-stage adaptive evaluation protocol.
- Opus-style CoT: Inherits Claude Opus's detailed chain-of-thought reasoning style, making it suitable for complex, multi-step problems.
- Bilingual Support: Optimized for English, with secondary support for Korean, Chinese, and Japanese.
- Robust Architecture: Utilizes a Qwen3.6 generation backbone with 64 layers and a context length inherited from its base, supporting long-chain reasoning.
Recommended Use Cases
- Graduate-level STEM Reasoning: Ideal for tasks like GPQA and science qualifying exams.
- Mathematical Problem Solving: Effective for MATH and AIME-style problems.
- Code Generation and Debugging: Suitable for HumanEval and MBPP tasks.
- Complex Multi-step Tasks: Excels in scenarios requiring detailed chain-of-thought processing.