Ricursive Intelligence releases Ripple-65B Preview

Ricursive Intelligence released Ripple-65B Preview on November 4, 2026, an open-weight, dense Transformer model with roughly 65 billion parameters. The model was pretrained on 12 trillion tokens and supports context lengths of up to 1 million tokens. It is licensed under Apache 2.0.

On the TerminalBench 2.1 agentic coding benchmark, Ripple scored 80.52, which Ricursive says is the highest among U.S. open-weight models, narrowly ahead of Reflection Beam 501B at 80.1 despite Ripple being far smaller. The company says full research and development, including pretraining, data work, evaluation, and posttraining, cost about $32 million, roughly one-tenth of what it estimates the previous open-weight leader cost. Ricursive reached its benchmark result using 1.4 million reinforcement-learning rollouts, about 70 times fewer than Reflection Beam, relying on synthetic data pipelines rather than purchased posttraining data.

The model uses a thin-and-deep architecture with a hidden dimension of 5120 across 80 layers, and Ricursive reports maintaining 40% model FLOPs utilization during training without custom kernels or quantization. Because of its smaller size, Ripple can run locally on hardware such as NVIDIA's DGX Spark. It is available now on Hugging Face, Together AI, GMI Cloud, and OpenRouter.

Why it matters

Ripple claims state-of-the-art agentic coding performance among U.S. open-weight models at a fraction of the typical development cost.

Source: Ricursive Intelligence