Thinking Machines ships RLVR results on Inkling 2 Small

The author compares running reinforcement learning with verifiable rewards on GLM-5.3-Flash using rented H200 nodes versus Inkling 2 Small using a serverless training API. On a fixed security-task exploit suite, 60 steps of GRPO on Inkling 2 Small outperformed the best GLM-5.3-Flash arm while significantly reducing costs, infrastructure bugs, and bring-up time. The author argues that open-weight U.S. models are essential for local security research and auditing. Open article →