Ricursive Intelligence has released Ripple-65B Preview, a dense language model with 65 billion parameters and Apache 2.0 weights. Available when checked on November 5, 2026, it gives developers a coding-focused model they can evaluate outside a hosted service. The launch post is undated, and we haven't run the model.
The lab reports a score of 80.52 on TerminalBench 2.1. That headline result comes with a qualification worth reading before choosing a coding agent: the model card explains that the surrounding evaluation software can change the score.
The agent harness affects the result
Ricursive's evaluation notes say its comparisons combine official model-card scores with internal evaluations where results were missing. It uses OpenCode when another harness isn't specified. The team also reports that switching to Terminus 2 with a 128K context and summarization generally lowered Terminal-Bench results in its comparisons.
Those disclosures make the score a starting point for testing. If you're considering the model for repository work, use the same agent software and task budget across candidates. Otherwise, a change in the runner can be mistaken for a change in model capability. We haven't independently verified Ricursive's benchmark results.
Open weights still leave deployment choices
The release includes BF16 and NVFP4 instruction-tuned weights, plus a BF16 base model. The main repository holds the BF16 instruction-tuned version. These are distinct artifacts, so an evaluation should record which one it uses.
In its architecture explanation, Ricursive says it chose a dense Transformer to reduce training and serving infrastructure complexity. That describes the lab's engineering choice; it doesn't establish the memory use or response speed you'll see on your hardware. We haven't measured either.
The next useful evidence is an independent run that records the checkpoint, numerical format, agent harness and task limits. Ricursive's own evaluation caveats give developers a practical checklist for that work, before treating its reported score as a reason to replace an existing coding setup.