The model was trained using a recipe inspired by that of deepseek-r1 [3], introducing self-reflection capabilities through reinforcement learning. Developed with NVIDIA tools, the company is releasing ...