Ricursive's Ripple-65B Claims Agentic Coding SOTA at $32M
Original: Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding
Why This Matters
Shows frontier-tier agentic coding performance is achievable at a fraction of typical pretraining costs.
Ricursive Intelligence released Ripple-65B Preview, a 65B dense open-weight model trained on 12T tokens for $32M, scoring 80.52 on TerminalBench 2.1 to claim U.S. open-weight SOTA on agentic coding, beating models up to 975B parameters.
Ricursive Intelligence, a full-stack AI lab focused on chip design, has released Ripple-65B Preview — a general-purpose variant of the proprietary model powering its AI chip design platform. The 65B dense Transformer model was pretrained on 12 trillion tokens with support for up to 1M context length, and achieves a TerminalBench 2.1 score of 80.52, outperforming Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8), and NVIDIA Nemotron 3 Ultra 550B (56.4).
The total R&D cost was $32M — roughly one-tenth what Ricursive estimates the prior open-weight SOTA cost to produce — covering everything from pretraining ladder development and data mixing to synthetic data generation and final training runs. Rather than purchasing posttraining data, the team built its own synthetic pipelines and quality filters.
For reinforcement learning, Ricursive used truncated importance sampling, score centering, and a linear length penalty, reaching 80.5 on TerminalBench with only 1.4 million rollouts — about 70x fewer than Reflection Beam. The team deliberately chose a dense architecture over Mixture-of-Experts, achieving 40% MFU throughout training with simple sharding and no custom kernels. The model is available on Hugging Face, Together AI, GMI Cloud, and OpenRouter.