RC RANDOM CHAOS

Ripple-65B: Open-Weight Agentic Coding SOTA for $32M

· via Hacker News

Original source

Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding

Hacker News →

Ricursive Intelligence released Ripple-65B Preview, a 65-billion parameter open-weight model trained on 12 trillion tokens that achieves state-of-the-art performance on agentic coding benchmarks. The model scores 80.52 on TerminalBench 2.1, outperforming significantly larger competitors: 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). It supports up to 1 million token context length and runs on Hugging Face, Together AI, GMI Cloud, OpenRouter, and Fireworks.

Development consumed $32 million in total R&D costs—approximately one-tenth the estimated budget of the prior open-weight agentic coding SOTA. Instead of mixture-of-experts scaling, the team built a dense Transformer with 80 layers and 5120 embedding dimension, co-designed for efficient execution on NVIDIA hardware. The architecture incorporates gated attention and value residuals. Using internally developed synthetic data pipelines rather than purchased training data, the model achieved its benchmark score through reinforcement learning with 1.4 million rollouts.

Iterative model ladder experimentation—testing architectural refinements at smaller scales before advancing to larger models—yielded 16x compute efficiency gains on code benchmarks at 4B scale compared to baseline approaches, suggesting these optimization patterns scale effectively at frontier model sizes.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.