FINAL-Bench/Darwin-4B-Opus

Darwin-4B-Opus: Reasoning-Enhanced MoE Model

Darwin-4B-Opus is a 7.9 billion parameter Mixture-of-Experts (MoE) model developed by VIDRAFT, leveraging the Gemma 4 Expert 4B architecture. It is distinguished by its use of the Darwin V6 engine, a novel merging approach that diagnoses parent models at the tensor level to assign independent optimal ratios, unlike conventional merging tools. This process integrates high-effort reasoning distillation from Claude 4.6 Opus, enhancing its capabilities in code, science, and analysis.

Key Capabilities & Features

• Advanced Merging: Employs Darwin V6 for diagnostic-guided evolutionary merging, analyzing tensor-level importance and functional impact.
• Reasoning Enhancement: Distills Claude Opus-level reasoning, particularly concentrated in FFN and final layers, as identified by the Model Diagnostic Scan (MDS).
• Efficient Architecture: Built on a Gemma 4 Expert 4B MoE base, offering strong performance in a highly efficient 4B parameter setup.
• Extensive Context & Language Support: Features a 128K context window and supports over 140 languages.
• Benchmark Performance: Achieves 82.92% on ARC-Challenge (zero-shot, loglikelihood).

Ideal Use Cases

• Resource-Constrained Environments: Designed for efficient deployment on single consumer GPUs, including NVIDIA RTX 4080/3090/4090 and T4.
• Edge Deployment & Rapid Prototyping: Its efficient 4B MoE architecture makes it suitable for applications requiring local inference.
• Reasoning-Intensive Tasks: Excels in scenarios demanding strong analytical, scientific, and coding reasoning, benefiting from Claude Opus distillation.
• Multilingual Applications: Broad language support enables diverse global use cases.