知Yizhi Daily
← Back to this edition

AI

Ripple-65B preview released with focus on low-cost agentic coding

Edition 2026-11-05 · 08:00 Beijing time

Ricursive Intelligence released the open-weight model Ripple-65B Preview. The model was trained from scratch, pretrained on 12 trillion tokens, supports a 1-million context, and scored 80.52 on TerminalBench 2.1. The company says its total R&D cost was $32 million and that it completed its reinforcement-learning leap with only 1.4 million rollouts.

A launch post on Hacker News says Ricursive Intelligence is a full-stack AI lab that co-optimizes hardware, software, and models to build self-improving systems. To support its chip design platform, the company trained a language model from scratch and released a general-purpose version, Ripple-65B Preview.

The model was pretrained on 12 trillion tokens and supports a context of up to 1 million. The article says Ripple-65B Preview scored 80.52 on TerminalBench 2.1, setting a new record for U.S. open-weight models in agentic coding and surpassing larger models such as Reflection Beam 501B, Meta Muse Spark Open 816B, Thinking Machines Inkling 975B, and NVIDIA Nemotron 3 Ultra 550B.

The company emphasized cost efficiency, saying the complete from-scratch R&D cost was $32 million, covering pretraining curriculum development, data mixtures, evaluation, debugging, synthetic data, post-training recipes, and the final pretraining and post-training runs. To its knowledge, this is about one-tenth the cost of previous open-weight state-of-the-art systems.

In terms of training methods, the team did not purchase post-training data. Instead, it built its own synthetic data pipeline, quality filtering, and coverage metrics, and used a hybrid reinforcement-learning approach involving truncated importance sampling, score centering, and a linear length penalty. The article says the model started from a strong SFT checkpoint and reached 80.5 with only about 1.4 million rollouts, roughly 70 times fewer than the previous state-of-the-art Reflection Beam. The model uses a modern dense Transformer architecture trained and optimized for serving on NVIDIA hardware. It is available on Hugging Face, Together AI, GMI Cloud, and OpenRouter, with Fireworks to follow.

Original sources

Hacker News ↗
Back to this editionOpen interactive reader