Ricursive Intelligence has introduced Ripple-65B Preview, an open language model aimed at coding tasks that use agents. The company reports a score of 80.52 on TerminalBench 2.1, which it describes as the highest result among open US models.
What the benchmark result says
TerminalBench 2.1 is the coding benchmark named in the company’s performance claim. Ripple’s reported score offers one measure of how it performs in that setting, though the source does not provide further benchmark results or a detailed comparison across models.
The article says other benchmark results put Ripple behind the newly released Inkling 2 Small. It also says Ripple is likely easier to deploy on users’ own hardware. That distinction matters for readers weighing performance against the practicalities of running a model themselves.
Ripple uses Terminus 2 in part and OpenCode in part as its harness. The source does not detail how the two contribute to the benchmark result, so the score should be understood in the context of the setup described rather than as a standalone measure of the model in every use.
A dense architecture and Nvidia focus
Ripple-65B Preview has a dense Transformer architecture with 65 billion parameters. The article contrasts that design with the more common mixture-of-experts approach, and reports that Ripple is optimized for Nvidia hardware.
These choices help describe the intended deployment profile. A dense model places its parameters in a single architecture, while the reported Nvidia optimization identifies the hardware ecosystem the model targets. The source does not give hardware requirements or provide a detailed account of performance on particular machines.
Training scale and reported cost
Ricursive says it pretrained Ripple on 12 trillion tokens. That figure describes the volume of training data processed, not a claim about how much information the model can reliably recall or how well it handles every coding task.
The company also puts total development and training costs at $32 million. That amount reportedly covers all research and development expenses, giving a broad picture of the resources behind the model. The source does not break down the total into individual cost categories.
For prospective users, the cost claim and the model’s hardware focus provide context about how Ripple was built. They do not by themselves establish what it would cost an outside team to run, adapt, or maintain the model.
Open access and practical considerations
Ripple-65B Preview is available on Hugging Face under the Apache 2.0 license. That makes its licensing and availability part of the announcement, alongside the benchmark result and architecture.
Anyone assessing the model can consider several points together:
- Its reported TerminalBench 2.1 score is 80.52.
- The company describes that as the best result among open US models.
- The model has 65 billion parameters and is optimized for Nvidia hardware.
- It is distributed under the Apache 2.0 license.
- The article says it trails Inkling 2 Small on other benchmark results, while likely being easier to deploy on users’ own hardware.
Those details frame Ripple as an open model with a strong reported result on one agentic coding benchmark, while leaving broader comparisons and deployment requirements for users to evaluate.