FINAL-Bench/Darwin-31B-Opus
Darwin-31B-Opus is a 31 billion parameter reasoning-enhanced language model developed by VIDRAFT, built on the Gemma 4 Dense architecture with a 32768 token context length. It was created using the Darwin V6 diagnostic-guided evolutionary merge engine, combining google/gemma-4-31B-it with TeichAI/gemma-4-31B-it-Claude-Opus-Distill. This model excels in complex reasoning tasks, achieving 85.9% on the GPQA Diamond benchmark with its Darwin-DELPHI test-time engine, and supports over 140 languages.
Loading preview...
Overview
Darwin-31B-Opus is a 31 billion parameter model developed by VIDRAFT, leveraging the Gemma 4 Dense architecture with a 32768 token context window. It is distinguished by its unique creation process using the Darwin V6 engine, which performs a diagnostic-guided evolutionary merge of two parent models: google/gemma-4-31B-it and TeichAI/gemma-4-31B-it-Claude-Opus-Distill.
Key Capabilities & Features
- Advanced Reasoning: Achieves 85.9% on the GPQA Diamond benchmark (a PhD-level science reasoning test) when paired with the Darwin-DELPHI test-time reasoning engine.
- Diagnostic-Guided Merging: The Darwin V6 engine diagnoses parent models at the tensor level, assigning independent optimal ratios to 1,188 tensors, a significant departure from conventional uniform merging methods.
- Multilingual Support: Capable of processing over 140 languages.
- Optimized for Reasoning: The merge process specifically favored the 'Mother' model (Claude-Opus-Distill) for its strong reasoning concentration in later layers, while retaining the 'Father' model's (Gemma 4) attention structure for multimodal and long-context capabilities.
- Chain-of-Thought: Supports
enable_thinking=Truefor chain-of-thought reasoning.
When to Use This Model
- For applications requiring high-level scientific or complex reasoning.
- When seeking a model with strong performance on challenging benchmarks like GPQA Diamond.
- For tasks benefiting from a long context window (32768 tokens) and multilingual capabilities.
- Developers interested in advanced model merging techniques and their practical applications.