An open model enters the coding race
VentureBeat reported on Ricursive Intelligence’s launch as a challenge to both proprietary coding systems and other open-weight models. The company’s pitch is not based only on scale. Ripple-65B Preview combines downloadable weights, local deployment and hosted access, giving organizations more control over where code and prompts are processed.
That control could matter as coding assistants move deeper into company infrastructure. A model that can run locally may be easier to integrate with private repositories, internal documentation and regulated development environments. It also gives engineering leaders an alternative to sending sensitive code to a third-party service.
In its launch announcement, Ricursive Intelligence describes Ripple-65B Preview as a dense model with 65 billion parameters, trained on 12 trillion tokens. The company says it is designed for coding and agentic computer use, and claims a U.S. open-weight lead on two major coding benchmarks. On TerminalBench 2.1, Ricursive reports a score of 80.52.
Those figures position Ripple as a serious contender in a crowded field that includes Chinese open-weight models and closed systems from larger technology companies. Yet the more consequential question is not whether Ripple can produce an impressive score. It is whether developers can trust it during the unglamorous hours of real work, when a task involves incomplete instructions, hidden dependencies, unfamiliar repositories and tools that occasionally fail.
Licensing as a product decision
The official Ripple-65B Preview model repository provides the downloadable weights in BF16 format under an Apache 2.0 license. That permissive license allows businesses to use, modify and distribute the model under broad conditions, making it easier to build commercial products around the weights.
For a company evaluating an AI coding system, licensing can be as important as capability. A permissive model can be fine-tuned for a particular programming environment, embedded into an internal tool or deployed on infrastructure selected by the customer. Teams can also avoid making their long-term workflow dependent on one hosted provider’s pricing, access rules or product roadmap.
Local availability does not eliminate operational complexity. Running a 65 billion parameter model requires substantial hardware and engineering expertise, especially when multiple developers need access at once. The model repository includes technical details, training information and benchmark tables, but also notes caveats involving evaluation harnesses and in-house scores. That makes independent testing essential for companies considering deployment.
Ricursive says its full research and development cost estimate for the model was $32 million. If that estimate is accurate, it offers a striking comparison with the resources typically associated with frontier AI development. It also raises a practical question: can a relatively efficient training effort produce a model that remains competitive across the wide variety of tasks found in software development?
Early tests show an uneven picture
The available independent evaluations are critical, and no opposing assessment was provided in the supplied sources. Plotly reported that its independent data-analysis evaluation saw Ripple pass 35 of 43 questions. Plotly described the model as relatively inexpensive but slow, and said it missed all five hidden-pattern questions. The evaluation also found that Ripple produced a materially incorrect table in one task.
That result paints a familiar picture of modern coding models. They can be useful collaborators, particularly when the task is explicit and the expected output is easy to verify. They can become less reliable when success depends on noticing an unstated pattern, checking an assumption or interpreting data rather than simply transforming it.
SMF Clearinghouse reported a weaker result on its Cold Iron coding evaluation. The organization said Ripple scored 93 of 157 tasks, with many failures attributed to provider timeouts and missed output caps. It concluded that the model’s agentic-coding state-of-the-art claim was not reproduced on that suite.
Those findings do not settle Ripple’s value. Timeouts and output limits may reflect the surrounding provider or test configuration rather than the model alone. They do, however, illustrate why benchmark numbers need context. A coding assistant lives inside a system of prompts, tools, latency limits, context windows and retries. A model that performs well in a controlled test may still frustrate developers if it is slow, verbose or brittle when tools behave unpredictably.
The real test is daily work
Ripple-65B Preview matters because it makes open coding AI feel less like a research artifact and more like infrastructure. A company could download the weights, inspect the license, test the system against its own repositories and decide how much control it wants over deployment.
The next stage will be less about declaring a universal winner than measuring where the model fits. Some teams may value local privacy and customization over peak performance. Others may prefer a closed service that is faster and easier to operate. Ripple’s opportunity is to prove that openness can be more than a philosophical advantage. It must become a practical one developers notice every day.
Maya Lindqvist is not a person. No notebook, no deadlines, no face behind the name — just a byline this newsroom publishes under. Here is the production line underneath it, because a name beside a portrait reads like a journalist, and this one is not one.
The models. Writing: gpt-5.6-luna and qwen3-max. Out on the live web: gpt-5.6-luna and gpt-5.6-terra. Pictures: gpt-image-1 and gpt-image-1-mini. Swap one in the newsroom and this line swaps with it — it is read off the machines, not typed here.
How a story is made
- Research. The searching model reads around the story, pointed at primary sources — the filing, the post, the repository — rather than at somebody else's write-up of them.
- Writing. The writing model drafts it against what was found, at Maya Lindqvist's usual length and in Maya Lindqvist's usual register.
- The loop. A reviewer reads the draft and sends it back with notes. Then reads it again. A piece can go round several times before it leaves the building.
- Enrichment. A quotation has to appear word for word on the page it is taken from. A chart may only use figures that appear in the source it cites. Whatever fails is dropped, and the reason is kept.
- Fact check. A last pass hunts for claims the article makes and its sources do not.
- A human stop. Sensitive subjects are held for a person to read before publication, and a person can kill any of it at any point.
If that sounds less like a newsroom and more like a factory: quite. It is called Press Factory.
This article was generated using AI and published automatically without human pre-publication review.
How this article was made
The article was produced by the Grandmonts Media News Engine using automated research, drafting and verification workflows. No human editor reviewed the article before publication. Grandmonts Media remains responsible for the published content. Errors can be reported at office@grandmonts.cz.