Ricursive Releases Ripple-65B Preview, Setting U.S. Open-Weight SOTA on Agentic Coding

• Ricursive Intelligence releases Ripple-65B Preview, a general-purpose open-weight model that achieves the U.S. open-weight state-of-the-art score of 80.52 on TerminalBench 2.1.
• The model was trained for a total R&D cost of $32M, which is one-tenth the estimated cost of the previous open-weight SOTA on agentic coding.
• Ripple-65B Preview is pretrained on 12 trillion tokens and supports up to 1M context length, utilizing a dense Transformer architecture optimized for NVIDIA hardware.
• The model surpasses significantly larger U.S. open-weight models, including Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8), and NVIDIA Nemotron 3 Ultra 550B (56.4).

Technical Approach

• Uses a thin-and-deep architecture with a narrow d_model of 5120 and 80 total layers, maintaining a 40% MFU without custom kernels or quantization.
• Achieves a TerminalBench 2.1 score of 80.5 with only 1.4M rollouts, roughly 70x fewer than the previous SOTA, by using a mixed RL recipe with truncated importance sampling and score centering.
• Employs a custom pretraining dataset with upsampled code, improving compute efficiency on code by up to 16x compared to the Olmo 3 baseline at 4B scale.

Availability

• Available on Hugging Face, Together AI, GMI Cloud, and OpenRouter, with Fireworks support launching later the same day.
• Designed to run efficiently on local devices such as NVIDIA DGX Spark due to its small parameter envelope.