A new contender in open-weight models
Ricursive Intelligence, an artificial intelligence lab founded by former Google researchers, presented Ripple-65B Preview on November 4, an open-weight model that the company describes as the best in the United States for performance in so-called agentic coding, that is, programming entrusted to autonomous systems capable of carrying out complex tasks. The news was reported by IT之家 on November 6. According to the company itself, this is an assessment that concerns exclusively U.S. open-weight models, a category that in recent months has seen strong competition among independent labs and large technology groups.
The model was created as a generalist variant of a technology previously developed by Ricursive Intelligence for AI-assisted chip design platforms. This transition from a specialized domain to a general-purpose model is the main novelty: a company that until now operated in a vertical sector enters the market for open foundation models, where competition is measured above all on reasoning capabilities and execution of real tasks.
Ripple-65B Preview is a dense model, pre-trained on 12 trillion tokens and with a context window of up to 1 million tokens. These are technical characteristics that place it in the large-model bracket, but with a non-sparse architecture, that is, without dynamically selected experts. The company claims that development and training from scratch cost $32 million, about 2.14 billion yuan at the exchange rate cited by the source.
The claimed comparison with larger models
According to Ricursive Intelligence, Ripple-65B Preview achieves better results on TerminalBench 2.1 than much larger competitors such as Meta Muse Spark Open 816B and Nvidia Nemotron 3 Ultra 550B. The data point is particularly relevant because it suggests that a more compact model can compete with systems that, at least in the number of parameters indicated in their names, exceed it by more than ten times. If confirmed, the result would indicate that training efficiency and data quality matter more than dimensional scale alone, a central theme in the debate on model development.
It should nevertheless be emphasized that the comparison is a statement by the company that developed the model and is not independently verified in the available material. TerminalBench 2.1 is a specific benchmark for terminal tasks and agentic programming, so the results do not automatically imply general superiority on other types of tasks, such as linguistic reasoning, multimodal understanding or production use. Moreover, it is unclear whether the models cited as competitors were evaluated with the same settings, the same number of attempts or the same inference conditions.
What remains to be verified
The material does not indicate whether Ripple-65B Preview is available for public download, under which license it is distributed, or what the commercial terms of use are—decisive elements for assessing the real impact of an open-weight model. Nor is it specified whether the $32 million cost includes computing infrastructure, training data or only certain phases of the process.
The news is of concrete interest to the sector: the entry of a lab specializing in chip design into the market for open general-purpose models adds a player to U.S. competition and offers a possible signal that compact models, if well trained, can reduce development costs compared with large proprietary systems. It remains to be seen, however, whether the claimed performance will be reproduced by independent evaluations and whether the model will actually be accessible to the developer community—two conditions necessary for the leadership claim to carry weight beyond the announcement.
