Ricursive Intelligence Releases Ripple-65B Preview for Agentic Coding

Overview

Ricursive Intelligence has introduced Ripple-65B Preview, a dense, open-weight language model designed for agentic coding tasks. The model, which was trained for $32 million, is reported to achieve a score of 80.52 on TerminalBench 2.1, positioning it as a competitive alternative to significantly larger open-weight models. The release aims to demonstrate the efficacy of co-optimizing hardware and software for specialized AI workloads, particularly in chip design.

Model Architecture and Training

Ripple-65B Preview utilizes a dense Transformer architecture rather than the Mixture-of-Experts (MoE) approach favored by many current frontier models. This design choice was intended to simplify infrastructure requirements while maintaining high compute efficiency. The model features a "thin-and-deep" aspect ratio, with a model dimension of 5120 and 80 layers.

Key architectural optimizations include:

  • Overprovisioned attention heads with adjusted Sliding Window Attention (SWA) configurations.
  • Integration of FFN activation clipping, gated attention, and global NoPE (No Position Embedding).
  • Use of value residuals to improve reasoning performance.

The team reported a 40% Model Flops Utilization (MFU) during training. Through iterative model laddering—a process of testing architectural interventions at smaller scales—the team improved compute efficiency on coding tasks by up to 16x compared to the Olmo 3 baseline.

Performance and Benchmarking

Ricursive Intelligence positions Ripple-65B as the U.S. open-weight state-of-the-art (SOTA) for agentic coding, citing a score of 80.52 on TerminalBench 2.1. The company claims this performance surpasses several larger models, including Reflection Beam 501B (80.1) and Meta Muse Spark Open 816B (76.4).

Community response to these benchmarks has been mixed. Some observers noted that the comparison set appears selective, pointing out the absence of high-performing models like the Qwen series in the provided charts. Others questioned the reliance on TerminalBench 2.1, suggesting that newer benchmarks might provide a more accurate reflection of current capabilities. Discussions on Hacker News highlighted that while the model's efficiency and training cost are notable, the "U.S. open-weight" designation is a specific, narrow category that does not account for the global landscape of open-weight models.

Deployment and Accessibility

Ripple-65B Preview is currently available on platforms including Hugging Face, Together AI, GMI Cloud, and OpenRouter. The model's dense architecture is intended to run efficiently on hardware with high compute FLOPs but lower memory bandwidth, such as NVIDIA DGX systems.

Community feedback indicates interest from users with high-memory local hardware, such as Mac Studios, who are looking for high-quality models that fit within their local memory constraints. However, some users noted that the dense 65B parameter count may limit its accessibility for hobbyists compared to smaller or MoE-based models that offer lower inference latency on consumer-grade hardware.

Sources

Related