Ricursive Intelligence released Ripple-65B Preview on November 4, adding a new open-weight option for developers building coding agents. The 65-billion-parameter dense model was trained from scratch, with post-training focused on agentic coding. Its Apache 2.0 weights are available on Hugging Face, alongside hosted access. Ricursive’s announcement describes the release; OpenRouter’s listing records the release date and API access.
What is available
- Downloadable weights: an instruction-tuned BF16 model, an experimental NVFP4 version and a separate base checkpoint.
- Long-context support: up to 1,048,576 tokens, with actual serving capacity dependent on deployment.
- Coding-agent performance: Ricursive reports 80.5 on Terminal-Bench 2.1. This is a vendor-reported result, not proof of universal superiority; the lab acknowledges that evaluation harnesses affect scores. See the model card for its evaluation details.
Why it matters
Builders gain another coding-focused model they can adapt and host themselves. For SunMarc, a useful evaluation would compare it on real website-maintenance tasks: correct changes, passing tests, latency and cost—not benchmark scores alone. Ripple remains a preview; we have not tested it.
For another recent approach to coding agents, see our coverage of Inkling 2 Small and its visual-feedback workflows.