Ricursive Intelligence releases Ripple-65B for agentic coding tasks

A new leader in coding

Ricursive Intelligence has launched Ripple-65B Preview, a 65-billion-parameter dense language model. The company claims it is currently the state-of-the-art US open-weight model for agentic coding tasks. These are tasks where an AI must navigate real environments, manage dependencies, and perform system administration.

The model was pretrained from scratch using 12 trillion tokens. It is designed to handle long contexts, supporting up to 1 million tokens. This allows the model to process and reason over very large amounts of data or code in a single session.

As OfficeChai reported, the model has shown strong performance on several benchmarks. On Terminal-Bench 2.1, it scored 80.5. This outperformed Reflection AI’s 501-billion-parameter Beam model, which scored 80.1. It also beat Meta’s 816-billion-parameter Muse Spark Open, which scored 63.8 on the same test.

Efficiency over sheer size

One of the most notable aspects of Ripple-65B Preview is its development cost. Ricursive Intelligence states the total R&D cost was around $32 million. They describe this as roughly a tenth of what the previous open-weight leader in agentic coding cost to build.

The company achieved these results using 1.4 million reinforcement learning rollouts. This is a fraction of the amount of data and compute that competing labs have reportedly used to reach similar performance levels. They also claim a 16x compute efficiency improvement on code compared to their own Olmo 3 baseline.

The lab attributes these gains to specific changes in scaling laws, model architecture, and data composition. By optimizing how the model learns, they were able to produce a smaller model that rivals systems many times its size. This suggests that raw scale is not the only path to high-level reasoning.

Dense vs. Sparse architecture

Ripple-65B Preview uses a dense architecture, which means it utilizes all 65 billion of its parameters for every single request. This stands in contrast to models like Beam, which use a sparse mixture-of-experts (MoE) approach. An MoE model only activates a small portion of its total parameters at any given time.

Ricursive Intelligence positions this dense design as a primary selling point. Because it doesn't rely on the massive infrastructure required by huge MoE models, Ripple is intended to fit on devices with high compute but limited memory bandwidth. Examples of such hardware include a MacBook or an NVIDIA DGX Spark.

This architectural choice makes the model more accessible for local deployment. It allows users to run a high-performing coding assistant without needing a massive data center. The goal is to provide power that fits within the constraints of professional workstation hardware.

The lab behind the model

Ricursive Intelligence is a Palo Alto-based lab founded in December 2025 by Anna Goldie and Azalia Mirhoseini. The founders were previously known for their work at Google Brain on using reinforcement learning for chip floorplanning. They launched the company with a $35 million seed round led by Sequoia Capital.

The lab quickly scaled, raising a $300 million Series A at a $4 billion valuation just two months after launch. Their core mission is to close the feedback loop between AI models and the chips that power them. They aim to use AI to compress semiconductor design cycles from years down to weeks or months.

Ripple-65B Preview is the lab's first public move into general-purpose language models. However, the two efforts are linked. A model that can reason about code well is also highly effective for the RTL generation and design-automation work that powers the company's core platform.

Availability and verification

The weights for Ripple-65B Preview are now available to the public under an Apache 2.0 license. Users can find the model on Hugging Face. Hosted access is also currently live through several providers, including Together AI, GMI Cloud, and OpenRouter.

While the performance numbers are impressive, it is important to note they are self-reported by the company. These results have not yet been independently verified by third parties. The lab is currently positioning the model as a tool to help US labs close the gap with Chinese open-weight models.

The model's focus is specifically on agentic coding rather than general-purpose capabilities. By mastering complex software engineering tasks, Ripple-65B Preview aims to provide a high-utility tool for developers. It represents a shift toward specialized, efficient models that can perform complex work on modest hardware.