FINAL-Bench/Darwin-31B-Opus

Hugging Face
VISIONConcurrent Unit Cost:2Model Size:31BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 6, 2026License:apache-2.0Architecture:Transformer0.1K Open Weights Featherless Exclusive Warm

FINAL-Bench/Darwin-31B-Opus is a 31 billion parameter language model developed by VIDRAFT, built on the Gemma 4 Dense architecture with a 256K context length. This model is enhanced for reasoning tasks through a diagnostic-guided evolutionary merge process, combining google/gemma-4-31B-it with a Claude-Opus-Distill model. It achieves 85.9% on the GPQA Diamond benchmark with its Darwin-DELPHI test-time engine and supports over 140 languages, making it suitable for complex analytical and scientific reasoning applications.

Loading preview...

Overview of Darwin-31B-Opus

Darwin-31B-Opus is a 31 billion parameter language model developed by VIDRAFT, utilizing the Gemma 4 Dense architecture. It is distinguished by its unique creation process: a diagnostic-guided evolutionary merge using the Darwin V6 engine. This engine intelligently combines google/gemma-4-31B-it (Father) and TeichAI/gemma-4-31B-it-Claude-Opus-Distill (Mother) by analyzing and assigning optimal ratios to 1,188 tensors individually, a significant departure from conventional uniform merging techniques.

Key Capabilities and Features

  • Advanced Reasoning: Achieves an impressive 85.9% on the GPQA Diamond benchmark (a PhD-level graduate science reasoning test) when used with the proprietary Darwin-DELPHI test-time reasoning engine.
  • Diagnostic-Guided Merging (Darwin V6): Employs a sophisticated Model Diagnostic Scan (MDS) to profile parent models at the tensor level, measuring entropy, standard deviation, L2 norm, and functional importance across 5 diagnostic probes (REASONING, CODE, MATH, KNOWLEDGE, LANGUAGE). This allows for highly optimized, per-tensor merging ratios.
  • Hybrid Architecture: Based on Gemma 4 Dense, it features Hybrid Attention (Sliding Window + Global) and supports a 256K context length.
  • Multilingual Support: Capable of processing over 140 languages.
  • Optimized for Reasoning: The merge process specifically favored the Mother model's (Claude-Opus-Distill) reasoning capabilities, particularly in FFN layers and the L50-L59 reasoning core, while preserving the Father model's attention structure for multimodal and long-context handling.
  • Thinking Mode: Supports enable_thinking=True for chain-of-thought reasoning.

Ideal Use Cases

  • Complex Scientific and Academic Reasoning: Its high performance on the GPQA Diamond benchmark makes it well-suited for tasks requiring deep analytical and scientific understanding.
  • Advanced Code and Analysis: Benefits from the Claude Opus distillation, enhancing its capabilities in code generation and complex data analysis.
  • Multilingual Applications: Broad language support makes it versatile for global applications requiring nuanced understanding across many languages.
  • Research and Development: Developers interested in advanced model merging techniques and fine-grained control over model capabilities will find the Darwin V6 methodology particularly insightful.

Finetunes
1 models