FINAL-Bench/Darwin-9B-Opus

VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 4, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Darwin-9B-Opus is a 9 billion parameter dense reasoning model developed by VIDRAFT, built on the Qwen3.5-9B architecture. It features a 131,072 native token context length and supports 201 languages, utilizing a unique diagnostic-guided evolutionary merge (Darwin V5) for enhanced reasoning capabilities. This model is specifically optimized for complex reasoning tasks, incorporating a `` tag for chain-of-thought processing.

Loading preview...

Darwin-9B-Opus: Diagnostic-Guided Evolutionary Merge

Darwin-9B-Opus is a 9 billion parameter dense reasoning model developed by VIDRAFT, leveraging the Qwen3.5-9B architecture. This model is distinguished by its creation using Darwin V5, an evolutionary merge engine that employs layer-level diagnostics to guide the merging process, rather than standard merge libraries. The 'Father' model is the original Qwen/Qwen3.5-9B, while the 'Mother' is a LoRA SFT of Qwen3.5-9B fine-tuned with text-only Claude 4.6 Opus reasoning chains.

Key Capabilities & Innovations

  • Diagnostic-Guided Merging: Darwin V5 implements a DARE-TIES merge directly via PyTorch, using Model MRI (Layer-level profiling of tensor health indicators like L2 norm, Shannon entropy, and standard deviation) to determine per-tensor merge ratios. This allows for a more nuanced and effective integration of parent models.
  • Evolutionary Search: The model's development involves a two-phase evolutionary search. Phase 1 uses a heuristic proxy over 200 steps, followed by Phase 2 which conducts 10 steps of actual tensor merges and real benchmark evaluations (ARC-Challenge) to select the best model.
  • Enhanced Reasoning: Designed for reasoning tasks, it incorporates a <think> tag for chain-of-thought processing.
  • Extensive Context & Language Support: Features a 131,072 native token context length and supports 201 languages.

What Makes This Different

Unlike standard merging techniques that use uniform ratios or human-set parameters, Darwin V5 employs a dynamic, per-tensor ratio derived from MRI diagnostics (70%) combined with an evolutionary genome (30%). This approach allows for precise control over the merge, including 'transplant' functionality where an entire parent tensor is used if its quality score is significantly higher or lower. This method provides detailed, layer-by-layer quality comparisons, enabling better diagnosis of potential issues.

Use Cases

  • Complex Reasoning Tasks: Ideal for applications requiring advanced logical deduction and problem-solving.
  • Multilingual Applications: Its support for 201 languages makes it suitable for global use cases.
  • Research & Development: Offers a unique approach to model merging, providing insights into tensor-level diagnostics and evolutionary optimization.

Finetunes
1 models