"Darwin-27B-Opus: Surpassing the Foundation Model Without Training" (huggingface.co)

On April 12, 2026, the AI community witnessed a breakthrough with the introduction of Darwin-27B-Opus, a 27-billion-parameter model that surpassed its foundation model without undergoing any traditional training or gradient updates. Achieving an impressive score of 86.9% on the GPQA Diamond benchmark, which assesses scientific reasoning across disciplines like physics and biology, this model ranked 5th globally on the HuggingFace leaderboard, outperforming not only its parent model, Qwen3.5-27B, but also larger models like GLM-5.1 and Qwen3.5-122B. The significance of Darwin-27B-Opus lies in its innovative approach to knowledge organization. Utilizing a technique called evolutionary crossbreeding, it combines Feed-Forward Network (FFN) layers from different pretrained models without altering attention layers, thus ensuring the model’s reasoning capability remains intact. By employing a Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to optimize layer blending ratios, Darwin effectively creates superior configurations from existing models, demonstrating that the AI field can harness existing knowledge more efficiently than through conventional methods requiring massive computational resources. This enables the potential for more collaborative model development and reduces compute costs, prompting a reevaluation of how AI models are built and refined.