FINAL-Bench/Darwin-9B-Opus

VISIONConcurrency Cost:1Model Size:9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 4, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Darwin-9B-Opus is a 9 billion parameter dense reasoning model developed by FINAL-Bench, built on the Qwen3.5-9B architecture. It features a 131,072 native token context length and supports 201 languages, excelling in chain-of-thought reasoning. This model is uniquely created using the Darwin V5 evolutionary merge process, which employs MRI-guided per-tensor diagnostics and an evolutionary search to optimize performance, making it suitable for complex analytical tasks.

Loading preview...

Overview

Darwin-9B-Opus is a 9 billion parameter dense reasoning model developed by FINAL-Bench, utilizing the Qwen3.5-9B architecture. It is distinguished by its creation through the Darwin V5 engine, an advanced evolutionary merge process that does not rely on external merge libraries like mergekit. Instead, Darwin V5 implements DARE-TIES directly via PyTorch tensor operations, incorporating MRI-guided per-layer ratios for precise model merging.

Key Capabilities & Unique Features

  • Advanced Merge Algorithm: Employs a custom DARE-TIES implementation with per-tensor diagnostic-guided ratios, combining MRI analysis (70%) and evolutionary genome ratios (30%).
  • Diagnostic-Guided Merging: Utilizes "Model MRI" for layer-level profiling of tensor health indicators (L2 norm, Shannon entropy, standard deviation) to determine optimal merge ratios.
  • Evolutionary Optimization: Features a two-phase evolutionary search (CMA-ES) over genome space, with a heuristic proxy phase followed by real benchmark evaluation (ARC-Challenge) to select the best model.
  • Reasoning Focus: Optimized for chain-of-thought reasoning, indicated by the inclusion of a <think> tag for structured thought processes.
  • Extensive Context and Language Support: Offers a 131,072 native token context length and supports 201 languages.

What Makes This Different

Unlike standard merging techniques, Darwin V5 provides:

  • Per-tensor Ratio Selection: Ratios are determined individually for each tensor based on MRI diagnosis, rather than a uniform ratio across all tensors.
  • Pre-merge Analysis: Comprehensive tensor-level profiling (norm, entropy, std) is conducted before merging.
  • Post-merge Validation: Includes layer-by-layer comparison of the merged child model against its parents to detect interference or function loss.
  • Transplant Support: Allows for direct tensor copying from one parent if a tensor's MRI ratio falls below 0.05 or above 0.95, avoiding interpolation.

Recommended Use Cases

  • Complex Reasoning Tasks: Ideal for applications requiring deep analytical capabilities and structured chain-of-thought processing.
  • Multilingual Applications: Suitable for use cases spanning a wide array of languages due to its 201-language support.
  • Research and Development: Offers insights into advanced model merging techniques and evolutionary optimization for LLMs.

Finetunes
1 models