Eric111/FINAL-Bench_Darwin-4B-Opus

VISIONConcurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 12, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Darwin-4B-Opus is a 7.9 billion parameter Mixture-of-Experts (MoE) causal language model developed by VIDRAFT, built upon the Gemma 4 Expert 4B architecture. Utilizing the Darwin V6 diagnostic-guided evolutionary merge engine, it integrates Claude Opus-level reasoning distillation, making it highly efficient for edge deployment and resource-constrained environments. This model excels in reasoning tasks, achieving 82.92% on ARC-Challenge, while supporting a 128K context length and over 140 languages.

Loading preview...

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 stands out due to its unique creation process using the Darwin V6 engine, which performs a diagnostic-guided evolutionary merge. This method analyzes parent models at the tensor level, assigning independent optimal ratios to each tensor, a significant departure from conventional merging tools.

Key Capabilities & Differentiators

  • Reasoning Enhancement: Achieves Claude Opus-level reasoning distillation, particularly strong in later layers, as evidenced by its 82.92% score on ARC-Challenge.
  • Efficient Architecture: As the smallest in the Darwin Opus family, its 4B parameter MoE design is optimized for edge deployment, rapid prototyping, and resource-constrained environments, running comfortably on a single consumer GPU (e.g., 8GB VRAM for BF16).
  • Advanced Merging (Darwin V6): Unlike standard merge tools, Darwin V6 conducts a Model Diagnostic Scan (MDS) on parent models, analyzing entropy, standard deviation, L2 norm, and functional importance via 5 diagnostic probes (REASONING, CODE, MATH, KNOWLEDGE, LANGUAGE) to determine precise per-tensor merge ratios. It also incorporates an evolutionary optimization algorithm (CMA-ES).
  • Multilingual & Contextual: Supports over 140 languages and features a substantial 128K context length.
  • Thinking Mode: Includes an enable_thinking=True chain-of-thought capability.

Ideal Use Cases

  • Resource-Constrained Applications: Its efficiency makes it suitable for deployment on consumer-grade hardware or mobile devices.
  • Reasoning-Intensive Tasks: Excels in applications requiring strong logical deduction and problem-solving.
  • Rapid Prototyping: Offers a powerful yet accessible model for quick development and iteration.
  • Multilingual Applications: Capable of handling diverse language requirements.

Finetunes
1 models