MENU

Ripple-65B Preview: Ricursive Intelligence’s Open Model Specs

스타트업레시피 South Korea
Overview
Ricursive Intelligence has released Ripple-65B Preview, an open-weight model with 65 billion parameters that outperforms larger US models in agent coding. Developed for $32 million, the model demonstrates high efficiency in training costs and performance relative to its size.
In Depth

Key Findings

Ripple-65B Preview has achieved a score of 80.52 on the Terminal-Bench 2.1 agent coding benchmark, the highest among US open-weight models. It surpassed significantly larger models, including Beam 501B (80.1), Meta’s Muse Spark Open with 816 billion parameters (76.4), and Inkling 975B (63.8). The model also demonstrated competitive performance in SWE-bench Pro v2.0 Hard and DeepSWE 1.1.

Technical Details

The model was pre-trained on 12 trillion tokens and supports a context length of up to 1 million tokens. Ricursive Intelligence reports a total research and development cost of $32 million, covering the entire pipeline from phased pre-training experiments, data adjustment, and synthetic data generation to final training. This cost is approximately one-tenth of the development costs associated with previous top-performing open models.

To reduce costs, the company focused on computational efficiency in SFT and reinforcement learning rather than raw compute power. Notably, the company did not purchase post-training data, instead utilizing a proprietary synthetic data pipeline with quality filtering and coverage metrics. The number of attempts required to reach the 80.5 score on Terminal-Bench 2.1 was 1.4 million, which is 1/70th the number of attempts used by Beam.

Background & Context

Ricursive Intelligence is a company specializing in recursive self-improvement for AI-driven semiconductor chip design. The development of Ripple-65B Preview comes amid a trend of escalating training costs in the industry, with OpenAI’s Sam Altman noting GPT-4 cost over $100 million and Anthropic’s Dario Amodei suggesting future pre-training costs could reach $100 billion by 2026–2027.

Significance & Outlook

Designed for efficiency on NVIDIA hardware, Ripple-65B Preview’s relatively small parameter size allows it to run in local environments as well as large-scale data centers. The model is available for free download via Hugging Face and is accessible through API services including Together AI, GMI Cloud, OpenRouter, and Fireworks.

While leading in coding, the model ranked third behind Muse Spark Open and Beam 501B in other benchmarks, specifically HMMT (February 2026), FORTRESS Adversarial, and the public version of AutomationBench.

Source: https://startuprecipe.co.kr/archives/tech/5840663

Get our weekly technology intelligence — free

Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.

Subscribe Free — Weekly Tech Intelligence

By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.

  • Your email and selected fields are used only to deliver the newsletter.
  • We never share your information with third parties.
  • You can unsubscribe anytime via the link in each email.

See our Privacy Policy for details.

Takes about a minute · Unsubscribe anytime

Published by Troy-Technical, an independent site run by one engineer with a career in materials development.
About the author / Contact info@troy-technical.jp
Let's share this post !

Author of this article

TOC