Ripple 65B Preview tops TerminalBench 2.1 among U.S. open models

Ricursive introduced the Ripple 65B Preview, an open 65-billion-parameter model that, according to the company, achieves the highest score among US open models on TerminalBench 2.1. The result was reported by The Decoder on November 6, 2026, based on information from the company itself.

In this article you will see:

• What Ripple 65B Preview is and why it got attention
• What the TerminalBench 2.1 result says
• How it compares to Inkling 2 Small
• What still needs to be confirmed before using in production
• Frequently asked questions about Ripple 65B Preview

What is Ripple 65B Preview

Ripple 65B Preview is an open language model from Ricursive, with 65 billion parameters and an Apache 2.0 license. The company's proposal is to offer an open model made in the United States, with a focus on programming and agent tasks. The Apache 2.0 license allows commercial use and modifications, which makes the model interesting for teams that want to avoid dependence on a closed API.

According to the news, the total research and development cost was $32 million. That number matters because it puts Ricursive in direct competition with larger labs that spend much higher amounts on training. Even so, Ripple 65B Preview does not try to win all benchmarks. The stated focus is performance on terminal and programming tasks, where agents need to execute commands, fix errors and complete multi-step tasks.

The model is available on Hugging Face, which makes local testing and comparisons with other open models easier. For developers using AI, this means one more open model option to run on own infrastructure, with control over data and cost.

What the TerminalBench 2.1 result says

The main point of the announcement is the performance on TerminalBench 2.1, a benchmark that measures agent capability on terminal tasks. The company says Ripple 65B Preview achieves the highest score among US open models on that benchmark. The result is relevant because terminal and command line are common scenarios for code agents, especially in development, debugging and automation workflows.

The news also reports that Ripple uses two different harnesses in the tests: Terminus 2 in part of the runs and OpenCode in another part. This matters because the harness can influence the result. A model can appear better or worse depending on how the agent is executed, which tools are available and how the prompt is built. For that reason, the score should be read as the result of a specific set of conditions, not as absolute proof of superiority.

TerminalBench 2.1 measures practical tasks, but it does not cover all uses. It does not evaluate, for example, conversational response quality, safety, multimodal reasoning or cost per token. A model can lead this benchmark and still not be the best choice for your project.

How it compares to Inkling 2 Small

The same article compares Ripple 65B Preview with Inkling 2 Small, another open model recently released. According to the report, Ripple trails Inkling 2 Small on other benchmarks. This creates a common scenario in today’s market: one model wins on a specific task, while another wins on broader tasks.

In practice, the choice between the two depends on the work. If the focus is terminal automation, command execution and completing development tasks, Ripple can be a strong option. If the focus is overall performance on broader benchmarks, Inkling 2 Small appears to have the edge according to the numbers published.

This kind of comparison also shows why a single benchmark should not decide model adoption. The ideal is to test in your own repository, with the team’s real tasks, and measure time, cost and success rate. For those using code agents, it is also worth comparing latency, token consumption and stability on long tasks.

What still needs to be confirmed

The Ripple 65B Preview numbers come from the company, not an independent lab. That does not invalidate the result, but it calls for caution. Before adopting the model, it is worth confirming the score from independent sources, the exact version of the model tested and the harness configuration.

It is also important to note that the cited result is among open models from the United States, not all models in the world. Closed models and open models from other countries may have different scores. The announcement wording is specific and should not be read as overall leadership.

Another point is model maturity. A preview can change quickly. License, weights and documentation may be updated. Before putting it into production, check the version on Hugging Face, the date of the last update and the hardware requirements.

To expand the comparison with other recent benchmarks, see VIALS AA benchmark for scientific image reasoning and OpenAI introduces BioBasicBench biology test with AI.