Ripple-65B records 80.5 TerminalBench 2.1 at 1.4M rollouts

Ricursive trained Ripple-65B via iterated SFT followed by mixed RL that incorporated truncated importance sampling, score centering, and linear length penalties. Synthetic data pipelines generated long-horizon environments without purchased post-training data; quality filters and coverage metrics drove task difficulty increases as model capability rose. The resulting checkpoint posted 80.5 on TerminalBench 2.1 while consuming only 1.4 million rollouts.

TerminalBench 2.1 measures agentic coding success across multi-step software engineering workflows. Reflection Beam previously held the open-weight lead but required approximately 98 million rollouts under comparable evaluation conditions. Ripple-65B therefore demonstrates a 70-fold reduction in rollout budget while exceeding the prior score. This efficiency stems from targeted RL task curricula rather than scale alone.

US open-weight labs have historically trailed closed models on agent benchmarks because of data and compute constraints. Ripple-65B closes part of that gap through recipe-level improvements instead of larger pretraining runs. Comparable efficiency patterns appear in the Llama-3.1 post-training ablations and the Qwen2.5-Coder RL stage, both of which prioritized synthetic trajectory filtering over raw volume.

Operational deployment now favors models that reach usable agent performance inside modest rollout budgets. Teams can iterate on internal environments without matching the compute outlays of prior SOTA runs. Continued addition of harder tasks during training suggests further gains remain available before saturation on the current benchmark version.