Ripple-65B Preview is an open-weight coding model released by Ricursive Intelligence. Its model card describes a 65-billion-parameter dense transformer trained on 12 trillion tokens, with support for up to one million tokens of context.
Why the architecture matters
Ricursive chose a dense architecture rather than a mixture-of-experts design. Total parameter counts therefore need context: a mixture-of-experts model can activate only a subset of its parameters for a particular token.
In the company’s comparison, Ripple scored 80.52 on Terminal-Bench 2.1. The model card explains that some competing scores come from official model cards while others use its own evaluations, and that the agent harness can change results. The figure is a result on a defined benchmark, not a promise that Ripple wins every coding task.
Weights and local deployment
The release includes instruction-tuned BF16 and experimental NVFP4 versions as well as a base model. The model card lists an Apache 2.0 license. Ricursive describes local use on hardware such as NVIDIA DGX Spark; that does not imply that every ordinary computer can run it.
Start with the job, the available hardware and comparable test settings. Downloadable weights give users more control over deployment, while speed and quality still need to be measured in the actual environment.
