Ripple-65B Preview: New Open-Weight SOTA Model for Agentic Coding

Ricursive Intelligence has released Ripple-65B Preview, a general-purpose variant of the proprietary language model powering its AI chip design platform. Pretrained on 12 trillion tokens with support for up to a 1M context length, the model sets a new U.S. open-weight state-of-the-art for agentic coding.

• Ripple-65B Preview achieves 80.52 on TerminalBench 2.1, outperforming significantly larger open-weight models like Reflection Beam and Meta Muse Spark.
• The entire R&D and training process cost only $32 million, dramatically undercutting frontier training expenses.
• Its dense architecture enables strong reasoning performance and efficient execution on local hardware like NVIDIA DGX.

Architecture and Compute Efficiency

Rather than opting for a complex Mixture-of-Experts (MoE) approach, the team utilized a modernized dense Transformer architecture. This choice reduced infrastructure complexity and maintained a solid 40% MFU throughout pretraining and posttraining without custom kernels or quantization. The model features a thin-and-deep design with 80 total layers and overprovisioned attention heads.

By iteratively refining model ladders starting from Olmo 3, the researchers significantly improved code compute efficiency by up to 16x at the 4B scale. They relied entirely on proprietary synthetic data pipelines and custom RL recipes rather than purchasing external posttraining datasets, achieving rapid convergence with drastically fewer rollouts.

Availability and Ecosystem Impact

Ripple-65B Preview is now available on major platforms including Hugging Face, Together AI, GMI Cloud, and OpenRouter. By demonstrating that cutting-edge reasoning models can be built efficiently on tighter budgets, Ricursive Intelligence provides a powerful new tool for developers working on autonomous coding agents.