Ricursive's Ripple-65B: Open Apache 2.0 Model Scores 80.5% on TerminalBench

If you want to run a coding agent on your own servers without sending source code to an external API, the selection of open models has remained limited. Ripple-65B Preview from Ricursive Intelligence expands that selection: the model is available on Hugging Face under the Apache 2.0 license, meaning it can be fine-tuned, integrated into products, and sold without having to navigate restrictive terms.

What We Know About the Model

Ricursive Intelligence is best known as a company that designs AI chips, and language models are a new line of business for it. Ripple-65B Preview was pretrained on 12 trillion tokens. According to the company, it scored 80.52% on TerminalBench 2.1 — the best result among open US models. Two harnesses were used in the evaluation: partly Terminus 2 and partly OpenCode. This is an important detail: performance on agentic benchmarks depends heavily not only on the model itself, but also on how the tool-calling loop is implemented.

The architecture is also unusual for today's market. Most recent large models use a mixture-of-experts design, in which only a subset of parameters processes each token. Ripple is a dense 65-billion-parameter transformer optimized for Nvidia hardware. It is easier to deploy and more predictable in its behavior, although at the same performance level it is generally more computationally expensive per token.

The stated cost of the entire development effort, including research and training, is $32 million. Compared with leading labs, that is a modest sum, although it cannot be independently verified.

Comparison With Inkling 2 Small

The full picture is less rosy. Inkling 2 Small was released at the same time and is noticeably stronger on other tests:

• SWE-Bench Pro v2.0 Hard: 79.1% for Ripple versus 86.3% for Inkling
• DeepSWE v1.1: 44.3% versus 72.1%
• MCP Atlas: 71.67% versus approximately 85%

The DeepSWE gap is particularly revealing: Ripple trails by almost 28 points. MCP Atlas, meanwhile, evaluates tool use through the MCP protocol, which is critical for agents. Ripple's lead therefore applies to a specific benchmark and a specific category — open US models — rather than the entire market.

According to The Decoder, Ripple's advantage is probably easier deployment on in-house hardware. But that is an assumption, not a measured fact.

What to Do in Practice

First, do not trust a single number. An 80.5% score on TerminalBench does not mean the model will handle your monorepo. Test it on your own tasks — bug fixes, migrations, and test writing — using the harness you actually run.

Second, if legal clarity and control over data matter to you, Apache 2.0 plus US provenance resolve many compliance concerns. For teams that cannot send code outside their environment, this is a significant argument.

Third, the word Preview in the name is there for a reason. This is an early version, and its results will probably change. It is worth benchmarking now, but there is no need to rush it into production.

Bottom line: Ripple-65B is an intriguing open candidate for locally hosted coding agents, with a clear license and a relatively inexpensive development history. Against Inkling 2 Small, however, it looks more like an option for teams that value ease of deployment over maximum quality. For now, all figures come from the company itself and require independent verification.