tl;dr: Ripple-65B Preview, a US open-weight model, hit 80.5 on TerminalBench 2.1 for agentic coding using only 1.4M RL rollouts—roughly 70x fewer than the prior SOTA (Reflection Beam). The team skipped brute-force RL scaling and purchased post-training data, instead building their own synthetic data pipelines and a mixed RL recipe featuring truncated importance sampling, score centering, and linear length penalty on long-horizon synthetic environments, with task difficulty scaled progressively as the model improved.