Ricursive Intelligence, a five-person Palo Alto lab, has released Ripple-65B Preview, an open-weight coding model trained from scratch on 12 trillion tokens. The lab’s launch post describes a 65-billion-parameter dense Transformer with a one-million-token context window and Apache 2.0 licensing. The model weights are available on Hugging Face, while the OpenRouter listing, Together AI, GMI Cloud, and Fireworks provide hosted access.
- Architecture: Dense Transformer with 80 layers and a model width of 5,120
- Training: 12 trillion predominantly English and code tokens
- Context: One million tokens in the release, with training stages reaching two million
- License: Apache 2.0
- Reported budget: $32 million for model-development compute
- Primary focus: Agentic coding, including shell use, file editing, and iterative debugging
A 65B model takes on much larger MoEs
Ricursive reports a score of 80.52 on Terminal-Bench 2.1. These evaluations test different aspects of software work, from operating a terminal to resolving repository issues under an automated tool harness.
Several compared models contain hundreds of billions more total parameters, but total parameter count requires context. Mixture-of-Experts models route each token through a subset of their weights, while a dense model activates every parameter for every token. Storage requirements, active compute, memory bandwidth, and serving complexity therefore matter alongside headline model size.
The dense route around MoE complexity
Mixture-of-Experts architectures dominate many recent frontier open-weight releases because sparse activation can expand model capacity without applying every weight to each token. Ricursive selected a dense design to simplify pretraining, post-training, sharding, and inference infrastructure.
Ricursive says its team grew from one person to five while sustaining roughly 40% Model FLOPs Utilization during pretraining and post-training with standard sharding, no custom kernels, and no quantization. Model FLOPs Utilization, or MFU, estimates how much of an accelerator’s theoretical peak compute performs model training work. The figure indicates efficient hardware use, although it does not measure model quality or serving speed.
A thin, deep architecture
Ripple-65B uses a relatively narrow model width of 5,120 across 80 layers. During development, Ricursive moved from an OLMo 3 and Llama-like baseline toward a Gemma 3n-like architecture, which the team says improved stability at scale.
The final design combines several techniques: