Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding

Ricursive Intelligence, a full-stack AI lab, has achieved a U.S. Open-Weight State-of-the-Art (SOTA) on Agentic Coding with their proprietary model, Ripple-65B Preview. This general-purpose variant of their AI chip design platform model was trained from scratch to power their chip design platform.

The model, pretrained on 12 trillion tokens with a support for up to 1M context length, outperformed larger US open-weight models such as Reflection Beam 501B, Meta Muse Spark Open 816B, Thinking Machines Inkling 975B, and NVIDIA Nemotron 3 Ultra 550B in agentic coding tasks. Specifically, it achieved a U.S. open-weight state-of-the-art score of 80.52 on TerminalBench 2.1, surpassing the previous SOTA.

Ricursive Intelligence's model is dense, providing strong reasoning performance with low-latency token generation, and has a small parameter envelope that enables efficient local device running. Ripple-65B Preview is available on platforms such as Hugging Face, Together AI, GMI Cloud, and OpenRouter.

The development and training of Ripple-65B Preview cost only $32 million, significantly less than the estimated $100 million spent by OpenAI for GPT-4 training in 2023, and the projected $10-100B for training frontier AI models by 2026-2027. Ripple-65B Preview was designed for efficient training and serving on NVIDIA hardware, with a thin-and-deep model aspect ratio to reduce infrastructure complexity while still enabling strong reasoning and agentic performance.