Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding
Points and comments are a snapshot, not live.
Ricursive releases Ripple-65B Preview, a 65B open-weight model scoring 80.52 on TerminalBench 2.1 for agentic coding.
Ricursive Intelligence trained Ripple-65B Preview from scratch for $32M, claiming U.S. open-weight SOTA on agentic coding. The model surpasses Reflection Beam 501B (80.1) and Meta Muse Spark Open 816B (76.4) on TerminalBench 2.1. It uses a dense 65B-parameter Transformer with 80 layers, 1M context, and techniques like gated attention and value residuals. Pretrained on 12T tokens, it reached its score with 1.4M RL rollouts (70x fewer than Reflection). An internal variant powers Ricursive's chip design platform.
What commenters are saying
Commenters criticized the narrow category definition (U.S. open-weight) as diluting 'SOTA,' noting stronger models like Qwen 3.8 Flash Next (84.3) were excluded. Some noted practical reasons for the framing: U.S. government agencies cannot use Chinese models. Others questioned comparisons, pointing out Reflection Beam 1.0 or Nemotron 3.5 Super were omitted. A few saw value for domestic deployment, while most dismissed the 0.42-point lead as unimpressive. The name similarity to Recursive AI drew jokes.