StudioRicursive Intelligence releases Ripple-65B-Preview, a new 65B dense model
Original title: Ricursive Intelligence Ripple-65B-Preview: New Dense Model!
Ricursive Intelligence released Ripple-65B-Preview, a 65B-parameter dense Transformer pretrained on 12T tokens with up to 1M context, under Apache 2.0. The model self-reports 80.5 on Terminal-Bench 2.1, above larger US open-weight models like Reflection Beam 501B (80.1) and Meta Muse Spark Open 816B (76.4). Starting from an Olmo 3 baseline, the team improved code compute efficiency by 16x, up to 55x for the full hero run, using a Gemma 3N-like architecture (d_model 5120, 80 layers) with 40% MFU, no custom kernels, and BF16 training; a 1-to-5-person team trained it from scratch, and the model fits on consumer hardware like NVIDIA DGX Spark, with NVFP4 and BF16 base checkpoints also released.
The release includes the full training recipe and comparisons against much larger MoE models, showing how a small team built strong agentic coding with a dense architecture.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.