• Sources: primary, primary, discussion
  • Summary: Ricursive says Ripple-65B Preview supports up to 1M context and reports 80.52 on TerminalBench 2.1, a vendor-run result. Its RL recipe used truncated importance sampling, score centering and a linear length penalty and reached that score with 1.4M rollouts after an SFT stage built on its own synthetic data. The thin-and-deep dense architecture uses 80 layers, gated attention, global NoPE and value residuals and held 40% MFU without custom kernels.
  • Why it matters: A dense 65B agentic coding model is small enough for local hardware such as DGX Spark and documents a reproducible post-training recipe.