What Ricursive announced
Ricursive Intelligence, a company active in AI chip design, has trained its own open language model. It is called Ripple-65B Preview and was pretrained on 12 trillion tokens, according to The Decoder. The company claims the model scored 80.52 points on the TerminalBench 2.1 coding benchmark, the highest result among open models of US origin.
The score is a company statement, not an independently verified result. The same applies to the cost figure: Ricursive says total research and development cost $32 million, including all development and training expenses. The model is distributed on Hugging Face under the Apache 2.0 license.
Architecture and testing method
Ripple-65B Preview uses a dense Transformer architecture with 65 billion parameters, rather than the more widespread mixture-of-experts approach, and is optimized for Nvidia hardware. For the TerminalBench 2.1 tests, the company used partly Terminus 2 and partly OpenCode as the harness, that is, the infrastructure that runs and evaluates the model’s tasks.
The choice of a dense model over a mixture-of-experts is technically significant because it affects how the model is run and distributed, but the available material does not clarify what concrete advantages or disadvantages it entails in terms of performance or hardware requirements. The Apache 2.0 license, by contrast, is a confirmed fact and indicates that the model can be used and redistributed under the terms of that license.
The comparison with Inkling 2 Small and the limitations
The competitive picture is mixed. According to The Decoder, on other benchmarks Ripple does not match Inkling 2 Small, another recently released Western open model. The claimed record therefore concerns a single agentic coding benchmark, not an overall comparison. The source adds that Ripple might be easier to install on users’ hardware, but this is a hypothetical assessment, not a measured data point.
Several elements remain to be verified: the TerminalBench 2.1 results do not appear to have been replicated by third parties, the details of the benchmarks on which Ripple would be inferior to Inkling 2 Small are not specified, and it is not stated what hardware requirements are needed to run a dense 65-billion-parameter model. The preview nature also suggests that the model may change before a final version. The comparison between the two open models therefore remains partial and tied to the statements of their respective companies.
