Ripple-65B Preview: U.S. Open‑Weight SOTA Tackles Agentic Coding
agents training
| Source: HN | Original article
Ripple-65B Preview sets U.S. open‑weight state‑of‑the‑art performance in agentic coding, built for $32 M—about one‑tenth the cost of its predecessor.
Ricursive Intelligence announced the launch of Ripple‑65B Preview on 4 November 2026, positioning it as the United States’ open‑weight state‑of‑the‑art model for agentic coding. The 65‑billion‑parameter dense Transformer scored 80.52 on TerminalBench 2.1, edging out larger domestic rivals such as Reflection Beam 501 B (80.1) and Meta’s Muse Spark Open 816 B (76.4). Ripple achieved the result with a “thin‑and‑deep” 80‑layer architecture, 1.4 million reinforcement‑learning rollouts – roughly 70 times fewer than the rollout budget of Reflection Beam – and a context window of up to one million tokens.
What makes the launch notable is the reported development cost: Ricursive Intelligence says the full R&D budget, from pre‑training ladder design to data mixing and evaluation, was US $32 million, about one‑tenth of the expense associated with the previous open‑weight coding SOTA. Training consumed 12 trillion tokens, suggesting a highly efficient data pipeline.
The model’s performance matters because agentic coding – autonomous software engineering that can invoke tools, write, debug and refactor code without human prompts – is a fast‑growing segment of the AI market. A cheaper, high‑performing open‑weight alternative could broaden access for startups and enterprises that cannot afford proprietary offerings from Meta, Tencent or other cloud providers. It also intensifies the race to deliver cost‑effective, self‑hosting solutions, echoing recent moves such as Revolut’s shift to self‑hosted GLM models to slash inference spend.
Going forward, the community will watch for independent benchmark verification, adoption in developer toolchains, and whether Ricursive Intelligence opens the model for broader community contribution. Further releases may expand the model’s context length or integrate it into meta‑learning platforms, potentially reshaping the economics of agentic coding across the Nordics and beyond.
Sources
Back to AIPULSEN