Reproduce “Introducing Ripple-65B Preview, a Frontier Open-Weight Model” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ricursive Intelligence

Ricursive Intelligence releases Ripple-65B Preview, a general-purpose variant of the proprietary language model that powers its AI chip design platform. The dense 65B model was trained from scratch on 12 trillion tokens with up to 1M context, and reaches 80.52 on TerminalBench 2.1, a new U.S. open-weight state of the art for agentic coding. The full R&D cost was $32M, which the team estimates is one-tenth that of the previous open-weight SOTA.

Key insights

  • A dense model avoided MoE infrastructure complexity and kept a 40% MFU through pretraining and posttraining with simple sharding, no custom kernels, and no quantization.
  • The final model ladder matches the Olmo 3 ladder's code loss with up to 16x less compute at 4B scale, and its steeper slope suggests the gain grows with scale.
  • The small parameter envelope lets the model run efficiently on local devices like NVIDIA DGX Spark, which have high FLOPs but low memory bandwidth.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.