Introducing Ripple-65B Preview, a frontier open-weight model

Welcome to the first Ricursive newsletter.

Today, we release Ripple-65B Preview, a general-purpose open-weight version of the in-house model powering our AI chip design platform.

Ripple-65B Preview achieves U.S. open-weight SOTA (state-of-the-art) on agentic coding (80.5 on TerminalBench 2.1), surpassing significantly larger U.S. open-weight models such as 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 model is pretrained from scratch on 12 trillion tokens with support for up to 1M context length. Its small parameter envelope enables it to run efficiently on local devices such as NVIDIA DGX Spark.

We developed and trained Ripple-65B Preview for only $32M, which is one-tenth the cost of the previous open-weight SOTA on agentic coding, as estimated to the best of our knowledge. That figure is our full R&D cost starting from scratch.

Rather than investing in a brute-force RL (reinforcement learning) climb, we iterated on our SFT (supervised fine-tuning) and RL recipes with a focus on compute efficiency, using our own synthetic data pipelines. We reached 80.5 on TerminalBench 2.1 with only 1.4M rollouts, roughly 70x fewer rollouts than Reflection Beam, the prior SOTA. In pretraining, careful model laddering experiments produced a final model ladder that achieves similar code performance to the Olmo 3 ladder with up to 16x less compute.

We built Ripple-65B Preview to begin closing the loop between models and the hardware that fuels them. We were fortunate to build on top of existing open-source contributions, including those from the amazing Olmo team at AI2 (Allen Institute for AI), Prime Intellect, RadixArk, and NVIDIA.

Ripple-65B Preview is available today under Apache 2.0. Download the weights on Hugging Face, or try it on Together AI, GMI Cloud, and OpenRouter, with Fireworks going live later today.

If you want to help spread the word, a reshare or comment goes a long way: X post · LinkedIn post

The Ricursive team

P.S. At Ricursive, we have command of the full stack from model to silicon. If you want to join us and change the trajectory of AI and silicon: ricursive.com/careers