Ricursive Intelligence released the Ripple-65B Preview open-weights model on November 4 with a stated training cost of 32 million dollars
Ricursive Intelligence, founded by former Google researchers, released Ripple-65B Preview, a dense model pretrained on 12 trillion tokens with a 1M context window that scores above the much larger Meta Muse Spark Open 816B on TerminalBench 2.1.
On November 4, Ricursive Intelligence, an AI lab founded by former Google researchers, released Ripple-65B Preview, claiming the strongest agentic coding performance among open-weight models in the United States. Ricursive Intelligence previously focused on AI-driven chip design platforms, and Ripple-65B Preview is the general-purpose variant of its specialized model.
Ripple-65B Preview is a dense model pretrained on 12 trillion tokens with support for a 1 million token context, and its performance on TerminalBench 2.1 exceeds that of much larger competitors including Meta Muse Spark Open 816B and NVIDIA Nemotron 3 Ultra 550B. Ricursive Intelligence said the model cost 32 million dollars to develop and train from scratch.
The team previously concentrated on chip design platforms, and its models were built for a specific domain. This general-purpose release extends the same training approach to general coding tasks. The evaluation conclusions are vendor-reported and no independent reproduction is yet available.