The AI lab Magic pre-trained a base model matching DeepSeek models using half the compute required to train the original GPT-3