Ricursive opens Ripple-65B Preview for agentic coding
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
A medium-sized open model delivered a competitive agentic-coding result, giving developers another option for their own tools without a closed API.
Ripple-65B Preview is a dense 65-billion-parameter model pre-trained from scratch on 12 trillion tokens with a context of up to one million tokens. Ricursive released base and instruction-tuned checkpoints under Apache 2.0, plus an experimental NVFP4 version for DGX Spark-class devices. In Ricursive's own measurements, the model scored 80.52 on Terminal-Bench 2.1, narrowly ahead of larger open US MoE models. It trails some competitors on SWE-bench Verified, SciCode, DeepSWE and MCP Atlas. The company estimates total research and development cost at $32 million and says reinforcement learning required 1.4 million rollouts. All performance and cost figures are developer-reported.
Sources 2