# Ricursive releases Ripple-65B preview, claiming top US open-weight agentic coding

> Ricursive says its 65B dense Ripple-65B preview is the best US open-weight model for agentic coding, scoring 80.5% on Terminal-Bench 2.1.

- Status: reported (One outlet, not independently verified yet.)
- Topic: Models
- Published: 2026-11-05T06:25:13+00:00
- Updated: 2026-11-05T06:25:13+00:00
- Web page: https://ai-news.ikangai.com/story/ricursive-releases-ripple-65b-preview-claiming-top-us-open-weight-agentic-coding/
- Seen in: Latent Space
- Change: capability (New capability)
- Applies to: Ripple-65B

## What changed

Ricursive released a preview of Ripple-65B, a 65B dense open-weight model under an Apache 2.0 license.

Ricursive has released a preview of Ripple-65B, a 65B dense model that the company says is the best US open-weight model for agentic coding. Its headline number is 80.5% on Terminal-Bench 2.1, about a point above Beam-501B's 80.1%. Beam-501B is a mixture-of-experts model that activates 23B parameters per token. The comparison pits a dense model, which uses all 65B parameters on every token, against a larger sparse model that uses far fewer. Ricursive's claim is that the dense design competes with bigger models on this benchmark. It also says the model was pretrained from scratch.

The company says Ripple-65B is ahead of several larger mixture-of-experts models: Muse Spark Open-816B, Beam-501B, Inkling-975B and Nemotron 3 Ultra-550B. The margin over Beam-501B is less than one point on a single benchmark, so the claim rests heavily on that score and on the 'best US open-weight' framing. Critics quoted in coverage argued the framing is narrow. One said the 'US-only SOTA' label hides the gap between Ripple-65B and Chinese models. Another noted that Ricursive's release omits comparisons with Qwen. A third pointed out that the model still trails the 284B V4 Flash.

Ricursive reports 12T training tokens, a 1M context window and an Apache 2.0 license. The license lets companies use the weights commercially with few restrictions. The architecture uses value residuals and sliding window attention. The company presents its ablations on what it calls a scaling ladder. In that series, a new architecture trained on the same data as Olmo3 gave about a 3x gain, and a new dataset on the same architecture gave about 4x. Those figures are the analysis's reading of Ricursive's ladder. They suggest where the company thinks its gains came from: the data work contributed more than the architecture did.

On cost, Ricursive cites a $32M budget that it says covers full R&D, including the scaling ladder, data work and the final training runs. That detail matters for reading the claim. The budget covers the whole program, not one training run, so it is not directly comparable to figures that count only the final run. Prime Intellect ran the pretraining clusters and cites 16x pretraining compute efficiency for the run.

## Why it matters

The Apache 2.0 license and the claimed coding results give companies another permissively licensed option. The benchmark claims come from the company itself, though, and some comparisons are missing.

## What you can do

Evaluate Ripple-65B against current open models on your own coding tasks before changing any production system.

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "Ricursive releases Ripple-65B preview, claiming top US open-weight agentic coding", 2026-11-05, https://ai-news.ikangai.com/story/ricursive-releases-ripple-65b-preview-claiming-top-us-open-weight-agentic-coding/
