Thinking Machines pre-trained Inkling 2 Small on an NVIDIA GB300 NVL72 cluster containing 72 Blackwell Ultra GPUs per system in 68 hours, followed by nine days of reinforcement learning