FINAL-Bench/Darwin-4B-David

Darwin-4B-David: The First Second-Generation Evolutionary Model

Darwin-4B-David is a 4.5 billion parameter model developed by VIDRAFT, representing the first "evolution of evolution" in the Darwin series. It is a second-generation model, created by merging the first-generation Darwin-4B-Opus (evolved from gemma-4-E4B-it) with DavidAU's DECKARD-Expresso-Universe. This recursive merging strategy, guided by the Darwin V6 engine and Model MRI, allows for precise layer-wise integration of parent model strengths.

Key Capabilities and Differentiators

• Recursive Evolution: Darwin-4B-David is the first model to demonstrate a systematic, two-generation evolutionary merge, proving the concept of "Merge = Evolve" applied recursively.
• Exceptional Reasoning Efficiency: Achieves 85.0% on the GPQA Diamond benchmark (graduate-level scientific reasoning) with only 4.5 billion parameters. This score surpasses the 31B-class gemma-4-31B (84.3%), highlighting its superior parameter efficiency.
• "Thinking Mode" Generative Evaluation: The significant performance gains are observed in generative evaluations using a "thinking mode" (chain-of-thought), which captures the model's stable reasoning capabilities through maj@8 sampling.
• MRI-Guided Merging: Utilizes Darwin V6's Model MRI to scan weight divergence across layers and assign independent weight ratios, allowing for targeted absorption of features like the Mother model's creativity/character and Korean language hotspot.
• Edge Deployment Ready: Built on the Gemma 4 E4B architecture, it maintains Apache 2.0 licensing and is optimized for deployment on devices like the Jetson Orin NX 16GB and consumer GPUs, requiring approximately 16GB VRAM for BF16 precision.
• Multimodal Preservation: Freezes the Father model's vision (150M) and audio (300M) encoders during evolution, retaining image, video, and audio input capabilities.

When to Use This Model

• Complex Reasoning Tasks: Ideal for applications requiring strong logical and scientific reasoning, especially where generative "thinking mode" outputs are beneficial.
• Resource-Constrained Environments: Its high performance at a 4.5B parameter count makes it suitable for deployment on edge devices, mobile phones, or systems with limited GPU resources.
• Multilingual Applications: With support for 140+ languages and specific enhancements from the Mother model, it can handle diverse linguistic tasks.
• Applications Requiring Creativity and Insight: The integration of DECKARD-Expresso-Universe contributes to enhanced creativity and insight generation.
• Open-Source Collaboration: Represents a successful example of community synergy, leveraging contributions from active developers like DavidAU.