# Ricursive releases Ripple-65B Preview, an open-weight model it says leads on agentic coding

published: 2026-11-04T20:05:31.359Z
topic: infrastructure
source: https://www.ricursive.com/ripple
url: https://aiagentnews.io/stories/ricursive-releases-ripple-65b-preview-an-open-weight-model-it-says-leads-on

Ricursive Intelligence has released Ripple-65B Preview, an open-weight model that the company says is the top U.S. open-weight model on agentic coding. It reports an 80.52 score on TerminalBench 2.1 and says full development cost $32 million.

Ricursive Intelligence released Ripple-65B Preview, a general-purpose version of the proprietary model behind its chip design platform. The model was pretrained on 12 trillion tokens and supports context lengths up to 1M tokens. The company reports a TerminalBench 2.1 score of 80.52, ahead of open-weight models it names, including Reflection Beam 501B at 80.1 and NVIDIA Nemotron 3 Ultra 550B at 56.4.

Ricursive chose a dense Transformer architecture over mixture-of-experts to keep training and serving simpler on NVIDIA hardware. It says the model held an MFU of 40% through pretraining and posttraining, using simple sharding with no custom kernels or quantization. The company also says the model's smaller parameter count lets it run on local devices such as NVIDIA DGX Spark.

Ricursive puts full R&D cost at $32 million, which it estimates is one-tenth the cost of the previous open-weight state of the art on agentic coding. It reached its score with about 1.4 million rollouts, roughly 70 times fewer than Reflection Beam. The model is available on Hugging Face, Together AI, GMI Cloud and OpenRouter, with Fireworks due later.

## Why it matters if you build agents

Developers building coding agents get another open-weight model to evaluate, particularly if they want to run one on local hardware. The benchmark figures and cost estimate come from Ricursive itself, so test the model on your own agent tasks before relying on those results. Check the long-context behavior at high token counts as well, since the 1M figure is the company's stated maximum.