Ripple-65B Preview: Open US Coding Model Under Apache 2.0

What happened

Ricursive Intelligence is releasing Ripple-65B Preview, a 65-billion-parameter open-weight language model that the company says posts the top score among open-source US models on the TerminalBench 2.1 coding benchmark. The model is distributed under the Apache 2.0 licence, according to The Decoder, which reported the launch.

The company puts the total research and development cost at $32 million. Results are mixed beyond that single benchmark. The Decoder reports that Ripple trails Inkling 2 Small, another newly released Western open model, on several other tests.

Why it matters

For organisations building internal tools, the launch adds another permissively licensed option for agentic coding, where a model carries out multi-step programming tasks rather than answering one-off prompts. Open-weight releases under Apache 2.0 let teams host, inspect and adapt a model within their own infrastructure, which can matter for data governance and for integrating coding assistance into existing engineering pipelines.

The mixed picture also matters. A leaderboard result shows how a model performs on one test design, while a digital transformation programme depends on how it performs across the specific workflows an organisation runs. Teams weighing up open models should treat vendor-reported benchmark wins as a starting shortlist, not a procurement decision.

By the numbers

• 65 billion parameters in the Ripple-65B Preview model
• $32 million total R&D cost, as stated by Ricursive Intelligence
• Apache 2.0 licence, covering the model's full open release

The Renascence take

The headline is about a leaderboard position, but the more useful signal is the split result: a model can lead one coding test and still trail a peer elsewhere. Most readers will skip past that detail, yet it is exactly the kind of variance that shapes customer and employee outcomes once a tool reaches daily use. The principle underneath is simple: a single headline number anchors decisions, and anchoring is how teams end up buying capability they never measured against their own work.

Do not adopt Ripple on the strength of its benchmark score. Pick three real developer tasks your teams repeat every week, run both candidate models against them, and track completion rate, rework and time to a usable result. Let that evidence, not the press release, decide what goes into your engineering stack.