Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding (www.ricursive.com)

Ricursive Intelligence has announced the release of Ripple-65B Preview, an innovative open-weight language model that sets a new U.S. state-of-the-art (SOTA) in agentic coding. This model, pretrained on 12 trillion tokens and designed for efficient deployment on NVIDIA hardware, achieves an impressive score of 80.52 on TerminalBench 2.1, outperforming larger models like Reflection Beam 501B and NVIDIA Nemotron 3 Ultra, all while being significantly more cost-effective at just $32 million for total development, compared to estimates of up to $1 billion for competing models. The significance of Ripple-65B lies in its architectural choices and training efficiency, which include a thin-and-deep Transformer model, proprietary synthetic data generation, and optimized reinforcement learning strategies. These elements contributed to a remarkable reduction in compute resources needed for training; Ripple-65B requires approximately 70 times fewer rollouts than prior SOTA models. By tightly integrating hardware and software developments, Ripple-65B aims to enhance the performance and accessibility of AI chip design and coding capabilities, representing a pivotal advancement for the broader AI/ML community.