“DeepSeek R1, a reasoning model, shares many overlapping training steps with DeepSeek V3.”