Ripple-65B Preview: Dense 65B Model for Agentic Coding

Ricursive Intelligence, a startup founded by former Google DeepMind researchers, has released an open preview of Ripple-65B, a dense 65B model for agentic coding that scored 80.52 on Terminal-Bench 2.1, outperforming MoE models with 501–975 billion parameters, with development costs of approximately $32 million.

Compiled by Sergey KostenchukPublished 2026-11-09Updated 2026-11-09

2026-11-09 Research Google

🤖 Ripple-65B Preview: A Model for Agentic Coding

Ricursive Intelligence, founded by former Google DeepMind researchers Anna Goldie and Azalia Mirhoseini, has released a preview of Ripple-65B — a dense transformer without MoE featuring a context window of up to 1 million tokens. The model scored 80.52 on Terminal-Bench 2.1, outperforming Reflection Beam and Muse Spark.

🌍 The 65B model caught up to MoE models with 500–975 billion parameters, costing approximately $32 million with around 1.4 million RL rollouts — about 70 times fewer than Reflection Beam. Computational efficiency is the path to SOTA without massive budgets, although the 0.42-point margin is statistically noisy, and the model lags behind on SWE-bench Verified and GPQA Diamond.

👤 The weights are available under Apache 2.0 and can be run locally on consumer hardware such as the NVIDIA DGX Spark. API access is available via Together AI, Fireworks, and GMI Cloud at 20/60 cents per million input/output tokens.

Source 1: https://www.ricursive.com/ripple Source 2: https://huggingface.co/ricursiveAI/Ripple-65B-Preview

Sources