FINAL-Bench/Darwin-28B-Opus

Hugging Face
VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 24, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

Darwin-28B-Opus is a 27.6 billion parameter reasoning model from FINAL-Bench, built on the Qwen3.6 generation architecture with hybrid linear/full attention. It is an evolutionary merge of Qwen3.6-27B and a Claude Opus-distilled variant, designed to inherit Opus-style chain-of-thought reasoning. This model excels in graduate-level STEM reasoning, mathematical problem-solving, and complex multi-step tasks, achieving 88.89% on the GPQA Diamond benchmark.

Loading preview...

Darwin-28B-Opus: Evolutionary Reasoning Model

Darwin-28B-Opus is a 27.6 billion parameter model developed by FINAL-Bench, representing the first reasoning model in the Darwin series based on the Qwen3.6 generation backbone. It is the result of a Darwin V7 evolutionary merge, combining the strong bilingual reasoning capabilities of Qwen3.6-27B with Claude Opus 4-style chain-of-thought distilled behavior.

Key Capabilities & Performance

  • Exceptional Reasoning: Achieves an impressive 88.89% on the GPQA Diamond graduate-level reasoning benchmark (198 PhD-level questions) using a 3-stage adaptive evaluation protocol. This performance slightly surpasses its larger sibling, Darwin-36B-Opus (88.4%), and significantly outperforms its Qwen3.5-generation counterpart, Darwin-27B-Opus (86.9%).
  • Advanced Architecture: Utilizes the Qwen3.6 generation architecture with hybrid linear/full attention and 27.6 billion parameters (BF16 precision).
  • Opus-Style Reasoning: Inherits a sophisticated reasoning style through distillation from Claude Opus, making it adept at complex, multi-step thought processes.

Recommended Use Cases

  • Graduate-level STEM Reasoning: Ideal for tasks requiring deep analytical and problem-solving skills, such as GPQA and science qualifying exams.
  • Mathematical Problem Solving: Strong performance on challenges like MATH and AIME-style problems.
  • Code Generation and Debugging: Suitable for tasks like HumanEval and MBPP.
  • Complex Multi-step CoT Tasks: Excels in scenarios demanding intricate chain-of-thought reasoning.
  • Bilingual Reasoning: Offers strong support for English, with secondary capabilities in Korean, Chinese, and Japanese.

Limitations

  • Requires substantial VRAM (approx. 55 GB for full inference in bfloat16), typically needing high-end GPUs like an A100-80GB.
  • Optimized primarily for English, with secondary support for other languages.
  • Opus-style reasoning traces can be verbose, requiring careful management of max_new_tokens.

Finetunes
2 models