HYPE OR FACT? · CHECKED 5 NOVEMBER 2026
Ripple-65B Preview: 'US open coding lead' on Terminal-Bench
OUR VERDICT
Partly factOur original verdict: “Fact or hype: partly fact. The weights and licence are real; the headline benchmark wins are Ricursive's own runs and haven't been independently reproduced. On some other tests it does not lead.”
THE CLAIM
“Ripple-65B Preview (65B, Apache 2.0, up to 1M context) scores 80.5 on Terminal-Bench 2.1 and leads US open models on that test while staying small enough to run locally.”
What’s actually true
- Weights are public on Hugging Face under Apache 2.0 (ricursiveAI/Ripple-65B-Preview); product page at ricursive.com/ripple.
- Ricursive reports 80.5 on Terminal-Bench 2.1 and a dense 65B / ~12T-token pretrain; those scores are the company's own runs.
- The company itself notes harness choice moves scores a lot, and several knowledge/coding rows still favour larger rivals.
Sources
Cost
Open weights (Apache 2.0); you pay for hosting/inference. Chip-design platform is separate and commercial.
USE IT IF
Teams that want a self-hostable US open coding model to try on real agent tasks.
SKIP IT IF
Anyone treating the Terminal-Bench lead as independently verified, or needing a managed SaaS with SLAs out of the box.
How we knowChecked on 5 November 2026. Verdicts: Fact = the claim holds up; Partly fact = real, but with caveats or overstated parts; Hype = the headline claim isn't supported. Products change quickly, so check the maker's site before you buy. Our methods & sources ↗