Darwin Model Achieves 88.89% Without Training

VIDRAFT launched the Darwin-28B-Opus, a model achieving 88.89% on GPQA Diamond, leveraging a novel method of recombining existing model weights without any gradient-based training.

This development could significantly reduce the computational cost and resources needed for developing state-of-the-art reasoning models, by avoiding traditional post-training processes.

This new approach can empower smaller research teams or organizations lacking substantial computing resources to produce advanced models competitively.

While promising, there is uncertainty about the scalability and generalizability of this approach across diverse AI applications and datasets.