# Ricursive releases open Ripple-65B coding model

> The 65-billion-parameter model is available under Apache 2.0. Ricursive reports a strong coding benchmark score, but results vary across tests.

Oossa · 2026-11-06 · https://oossa.com/en/ricursive-releases-open-ripple-65b-coding-model

Event date: 2026-11-06

Ricursive Intelligence has released Ripple-65B Preview, an open-weight language model aimed at coding tasks that use tools and take multiple steps. The 65-billion-parameter model is available on Hugging Face under the Apache 2.0 license, which allows commercial use subject to the license’s terms. The Decoder reported the release on November 6, 2026.

Ricursive says it trained the model from scratch on 12 trillion tokens, mostly English text and code. It uses a dense Transformer architecture rather than a mixture-of-experts design, in which only parts of a model are activated for each task. The company says that choice helped keep training and serving infrastructure simpler.

## How did it score?

Ricursive reports a score of 80.52 on Terminal-Bench 2.1, a test of AI agents completing tasks in a computer terminal. Its model card lists a score of 80.5 for Ripple in both its self-reported results and a comparison using the Terminus 2 test setup. In that Terminus 2 comparison, it lists Reflection Beam 501B at 70.8, Muse Spark Open 816B at 78.7, and Inkling 975B at 56.2. These are company-reported figures, not results from an independent evaluation.

The model card combines results taken from other model makers’ published model cards with Ricursive’s own tests where results were missing. It also warns that scores can shift depending on the testing setup. And Ripple does not lead across every coding test: Ricursive lists its DeepSWE 1.1 score as 44.3, below Muse Spark Open’s 60.0, and its MCP Atlas score as 71.67, below Muse Spark Open’s 89.2.

## What can people download?

The preview includes an instruction-tuned model in BF16 and NVFP4 formats, plus a BF16 base model. The base model is intended for further training rather than direct instruction-following. Ricursive says Ripple supports a context window of up to 1,048,576 tokens, though the company’s page does not establish what hardware is needed to use that full capacity.

Ricursive says Ripple can fit on hardware such as its NVIDIA DGX Spark and provides an NVFP4 checkpoint for local development. That is a specific hardware claim, not evidence that the model will run comfortably on an ordinary laptop. The Decoder reports that Ricursive put total development and training costs at $32 million, including research and development.

## The facts

- Ripple-65B Preview has 65 billion parameters and was pretrained on 12 trillion tokens.
- Ricursive reports a score of 80.52 on Terminal-Bench 2.1.
- The model is available on Hugging Face under the Apache 2.0 license.
- Ricursive lists BF16 and NVFP4 instruction-tuned versions, and a BF16 base model.

## Why it matters

For developers who want to adapt or run an open model, Ripple offers an Apache-licensed option with downloadable checkpoints. Ricursive’s hardware claims point to local use on specific NVIDIA equipment, but the company has not established that the full model or its longest context will work on typical personal computers.

## Sources & references

1. [Ripple-65B Preview delivers strong agentic coding results as a fully Apache 2.0 open US model](https://the-decoder.com/ripple-65b-preview-delivers-strong-agentic-coding-results-as-a-fully-apache-2-0-open-us-model/) – The Decoder, 2026-11-06
2. [Ricursive Intelligence](https://huggingface.co/ricursiveAI/Ripple-65B-Preview) – Ricursive Intelligence

Last updated: 2026-11-06