Ripple-65B: A 65B Open Coding Model Claims to Beat Models 15x Bigger
In short
- Ricursive released Ripple-65B Preview on 4 November: a 65B open-weight coding model with a 1M-token context.
- It scores 80.52 on TerminalBench 2.1 by the company’s own test, ahead of open models 8 to 15 times its size.
- At 4 bits it needs about 33 GB of memory, against hundreds of GB for the big ones.
A new open-weight coding model just claimed the top spot among American open models. Ripple-65B Preview, from a company called Ricursive, scores 80.52 on TerminalBench 2.1. The company says that makes it the U.S. open-weight state of the art. It is a 65-billion-parameter model, small enough to run on one serious machine. That is the interesting part.
What happened
Ricursive published Ripple-65B Preview on 4 November. It is a dense transformer with 80 layers, trained on 12 trillion tokens, with a context window of up to 1 million tokens. The weights are on Hugging Face. You can also use it through Together AI, GMI Cloud and OpenRouter, and Fireworks was said to go live later on launch day. The company says training and R&D cost $32 million in total.
The headline number is TerminalBench 2.1, a test of agentic coding in the terminal. Ripple scores 80.52. The company lists Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8 and NVIDIA Nemotron 3 Ultra 550B at 56.4. Look at the sizes. Ripple is 65B and the others are 501B to 975B.
Two warnings. These are the company’s own results, not an independent test. And the launch page names no licence and gives no API price, so check both on the model card and with your host before you build anything.
How to use it
Try it through OpenRouter or Together first. Give it three real tasks from your own repo: a bug fix, a small feature and a test. Compare the result with what you use today. Only then think about running it locally. The page says it runs efficiently on devices like the NVIDIA DGX Spark, but gives no setup steps.
Do the math: 65 billion parameters at 4 bits is about 33 GB (65B x 0.5 byte, my estimate). At 8 bits it is about 65 GB. Compare that with the 500B to 975B models in the same table, which need hundreds of gigabytes. Now money: a top coding plan at $200 a month is $2,400 a year, about €2,200. A local model has no monthly fee, but you pay for the hardware once. Check current hardware prices before you decide.
What it means
If the claim holds, size is no longer the whole story for coding agents. A model one tenth the size of its rivals matching them means the cost of a good coding agent keeps falling. Closed labs lose some of their pricing power. Small teams win.
The word to watch is “Preview”. Wait for independent tests before you move production work.
Ivan’s Take
This is the kind of news I like: a small model doing a big model’s job. But it is a Preview, scored by its own maker, with no licence or price on the page. I would test it on three real tasks through an API this week and trust nothing until someone else repeats the numbers.
Source: Ricursive
If it doesn’t save you time or money, you won’t see it here
- Facts, not hypeEvery price and number is checked against the official source
- Always the real priceYou see what it costs before you click
- I say when it’s weakEven when it’s OpenAI or Google
