Open Model Ripple-65B Sets New Record Among US Open Models

Ricursive Intelligence, a company that previously designed chips for artificial intelligence, has released an open language model called Ripple-65B Preview. It has 65 billion parameters, a measure of a model's size. It was trained on 12 trillion tokens, the text used to teach it. On TerminalBench 2.1, a test of how well a model handles coding tasks from a command line, it scored 80.52 points. That is the best score among open-weight models built in the United States, the company said. Development and training together cost $32 million, according to Ricursive.

On other tests, Ripple-65B Preview falls behind. On SWE-Bench Pro v2.0 Hard, a benchmark for fixing real software bugs, it scored 79.1 percent. Inkling 2 Small, a rival open model from Thinking Machines, scored 86.3 percent on the same test. Ripple-65B Preview also uses a simpler design than many rivals: one dense network, not a mixture-of-experts system that splits work among smaller specialized parts. That design is built for Nvidia's chips and can be easier to run on a company's own hardware instead of through a cloud service. For companies that value that over topping a leaderboard, price and ease of running it may matter more than rank. The model is posted on Hugging Face, an online library for AI models, under the Apache 2.0 license, which allows free use and modification.