Ricursive Intelligence announced the release of Ripple-65B Preview on its official technical portal, presenting an open-weight 65-billion parameter dense foundation model trained from scratch on 12 trillion tokens.
The model achieved a benchmark score of 80.52 on TerminalBench 2.1, establishing a new U.S. open-weight state-of-the-art in autonomous terminal-based coding. In empirical evaluations, Ripple-65B Preview outperformed significantly larger mixture-of-experts architectures, including Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8, and NVIDIA Nemotron 3 Ultra 550B at 56.4.
A central achievement highlighted by Ricursive is capital efficiency: the lab developed and trained the model for a total research and development budget of $32 million. That sum encompassed pretraining ladder exploratory runs, synthetic generation, evaluation harnesses, post-training recipes, and final cluster executions. By comparison, frontier pretraining efforts across the industry frequently exceed several hundred million dollars.
Architecturally, Ricursive eschewed mixture-of-experts routing in favor of an 80-layer modernized dense Transformer featuring a narrow d_model of 5120. The design incorporates overprovisioned attention heads, an increased sliding window attention to global layer ratio, sigmoid gated attention, global NoPE positional representations, and value residuals. The engineering team sustained a 40 percent Model Flops Utilization across pretraining without custom assembly kernels.
In post-training, the team applied mixed reinforcement learning combining truncated importance sampling, score centering, and linear length penalties over high-quality synthetic environments. The model ascended to top-tier benchmark performance within 1.4 million rollouts—approximately 70 times fewer than the training budget demanded by Reflection Beam. Serving is supported across major providers at $0.20 input and $0.60 output per million tokens, though performance on non-coding reasoning domains remains uncharacterized in independent audits.