Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding

Ricursive Intelligence, a lab that builds AI for chip design, has released Ripple-65B Preview as an open-weight model and claims it sets a new U.S. open-weight high on agentic coding. Its headline number is 80.52 on TerminalBench 2.1, ahead of Reflection AI's Beam 501B at 80.1, Meta's Muse Spark Open 816B at 76.4, Thinking Machines' Inkling 975B at 63.8 and NVIDIA's Nemotron 3 Ultra 550B at 56.4.

The claim stands out for size and cost more than for the margin. The rivals' names carry parameter counts between 501 and 975 billion, while Ripple is a dense model, meaning it uses all of its weights for every token, which Ricursive chose over the mixture-of-experts designs other labs use in order to reduce training infrastructure complexity. Ricursive says its full research and development cost was $32 million, including experiments and the final training runs, and estimates that at one-tenth the cost of the previous open-weight leader on agentic coding, based on reports of Reflection's compute spending. It also says it reached its score with 1.4 million reinforcement-learning rollouts, roughly 70 times fewer than Reflection Beam, and that the model's small size lets it run on local hardware such as NVIDIA's DGX Spark.

All of these figures are Ricursive's own. The announcement does not say how the benchmark was run or under what settings, and a 0.4-point lead over Reflection Beam on a single benchmark is narrow. The model is labeled a preview and is a general-purpose variant of the proprietary model behind Ricursive's chip-design platform. The weights are on Hugging Face, and Together AI, GMI Cloud and OpenRouter serve the model; the announcement does not state a license.