Ricursive Releases Ripple-65B Preview, 65B Open-Weight Model

Ricursive Intelligence has released Ripple-65B Preview, a 65B dense open-weight language model, according to MarkTechPost. The release matters because the model is reported to score 80.5 on Terminal-Bench 2.1, putting a mid-sized open release ahead of larger US open models on that benchmark.

The name itself gives the most concrete details available so far. Ripple-65B Preview is described as a dense model with 65 billion parameters. It is a preview release, not presented in the report as a final production version, and it is made available as open-weight under the Apache 2.0 licence. Those three points — size, dense design, preview status and licence — frame what readers can and cannot conclude from the announcement alone.

What is Ripple-65B Preview?

According to MarkTechPost, Ripple-65B Preview is a 65B dense open-weight large language model from Ricursive Intelligence. The report does not provide additional architectural details, training data, context length, quantisation options or deployment requirements, so those questions remain open for readers who want to reproduce or run the model themselves.

The word dense is worth unpacking in general terms. A dense model uses most of its parameters for most tokens, as distinct from a mixture-of-experts design that activates only part of the network for a given token. That distinction affects memory use, compute cost and serving behaviour, but the report does not give Ripple-65B-specific figures for any of those measures. Readers should therefore treat the dense label as a design category, not as a performance guarantee.

Preview also matters. It signals that the release is intended for evaluation and feedback before any later, more polished version. For teams considering adoption, the practical question is not only what the model can do today, but what support, documentation and release cadence will follow. The announcement, as reported, does not answer those questions.

What does a score of 80.5 on Terminal-Bench 2.1 mean?

MarkTechPost reports that Ripple-65B Preview scores 80.5 on Terminal-Bench 2.1. Terminal-Bench is a benchmark centred on terminal-based tasks, which makes it relevant to coding, system administration and agentic workflows where a model must operate through a command line rather than simply answer a question in prose.

A single benchmark score should be read narrowly. It indicates performance on the specific tasks and scoring method used by Terminal-Bench 2.1, under the conditions used for this evaluation. It does not by itself establish how the model performs on other coding benchmarks, long-form reasoning, mathematics, multilingual tasks or everyday chat. Nor does it reveal cost, latency or reliability in a real deployment.

For developers, the useful next step is to ask how the score was produced and whether it can be reproduced. That includes the evaluation setup, the harness or agent framework used, any scaffolding around the model, and whether other models were evaluated under the same conditions. The report as summarised does not supply those details, so the 80.5 figure should be treated as a reported result awaiting independent checking rather than as a settled ranking.

What does open-weight under Apache 2.0 allow?

Ripple-65B Preview is reported to be released under Apache 2.0. In practical terms, an open-weight release means the model weights are made available for others to download, inspect, run and, subject to the licence terms, modify and redistribute. That is different from a model that is only accessible through a hosted API, where the weights remain closed.

Apache 2.0 is a permissive licence widely used for open-source software. For a model release, its importance is that it gives users clearer legal permission to use the weights in commercial and non-commercial projects, with fewer restrictions than a more limited research-only licence would impose. Readers should still check the actual release terms and any accompanying conditions before building on the model, because a summary in a news report is not a substitute for the licence text.

Open weights also change the evaluation dynamic. Independent researchers and developers can test the model on their own tasks, compare it with alternatives on equal footing, and report failures as well as successes. That openness is part of why the reported Terminal-Bench result will attract attention: the claim is, in principle, checkable by others rather than dependent on a closed system.

How does it compare with larger US open models?

According to MarkTechPost, Ripple-65B Preview scores ahead of US open models up to 15 times its size on Terminal-Bench 2.1. That comparison is the headline claim of the release. A 65B model outperforming models described as up to 15 times larger would, if confirmed under comparable conditions, suggest that parameter count alone does not determine terminal-task performance.

The comparison also needs careful handling. The report does not name the larger models, give their scores, or specify whether they were tested with the same agent setup, prompts, tools or time limits. Benchmark comparisons can shift substantially when the surrounding harness changes, especially for agentic tasks where the model interacts with an environment over multiple steps.

Readers should therefore separate two questions. First, what did the reported evaluation find? On the account given, Ripple-65B Preview scored 80.5 and placed ahead of the larger US open models mentioned. Second, how much weight should that carry in a buying or building decision? That depends on reproducibility, the relevance of terminal tasks to the reader's own work, and performance on the other dimensions that matter in practice, including speed, cost and robustness.

What should developers check before adopting it?

Because this is a preview, the sensible posture is evaluation rather than immediate adoption. Teams should first confirm where the weights and supporting files are published, what licence text applies, and what hardware or serving stack is required to run a 65B dense model at an acceptable speed. A model that fits a benchmark table may still be impractical if it needs infrastructure a team does not have.

Next, teams should test the model on tasks drawn from their own work. For coding and terminal use, that means real repositories, real shell commands and real failure modes, not only public benchmark items. It is also worth checking how the model behaves when it is uncertain, how it recovers from errors, and whether its outputs are easy to verify. Those qualities rarely appear in a single score.

Finally, teams should ask about maintenance. A preview release raises questions about bug fixes, security issues, documentation, community support and whether a later version will replace it. None of those points is answered by the headline score. Treating the release as a candidate for structured testing, rather than as a proven drop-in replacement, is the most defensible reading of the information available.

How can you explore this kind of work in Grafy?

The most direct Grafy connection is coding and terminal-style work. Grafy Code is an AI coding studio where developers can work with code in a development environment, which makes it a natural place to think about how a model that performs well on terminal tasks might be used in practice, from writing scripts to debugging and running code. Grafy does not claim to host Ripple-65B Preview, and this article makes no such claim; the point is to give readers a workspace for experimenting with coding workflows and comparing model behaviour.

For readers who want to study the announcement and related research more closely, Reader can be used to read, search and question PDFs and papers with AI, while Grafy Chat allows chat with many AI models side by side. Those tools fit the evaluation stage described above: checking claims, comparing explanations and designing tests before committing to a new model.

Frequently asked questions

Is Ripple-65B Preview released as an open-weight model?

Yes. According to MarkTechPost, Ricursive Intelligence released Ripple-65B Preview as a 65B dense open-weight language model. Open-weight means the model weights are made available for others to download and run, rather than being accessible only through a closed hosted service. Readers should check the original release for the exact files and terms.

Does Ripple-65B Preview score 80.5 on Terminal-Bench 2.1?

Yes, according to MarkTechPost, Ripple-65B Preview scores 80.5 on Terminal-Bench 2.1. That is the figure reported for the release. The report summarised here does not provide the full evaluation setup, so readers who want to verify the result should look for the underlying evaluation details and test the model on their own terminal tasks.

Is Ripple-65B Preview licensed under Apache 2.0?

Yes. According to MarkTechPost, Ripple-65B Preview is released under Apache 2.0. Apache 2.0 is a permissive open-source licence that generally allows use, modification and redistribution subject to its terms. Teams should read the licence text in the release itself before using the model in a product or service.

Source: MarkTechPost

---

Source: MarkTechPost — Ricursive AI Releases Ripple-65B Preview: A 65B Dense Open-Weight Model Scoring 80.5 on Terminal-Bench 2.1