Ripple-65B tops US open models on agentic coding at a $32M cost

Ricursive Intelligence, a lab that makes chip design tools, released Ripple-65B Preview, a model with open weights (anyone can download it). The company says it scores 80.52 on TerminalBench 2.1, a test of coding agents. Reflection Beam 501B scores 80.1. The company says its full R&D cost was $32M. These are the company results.

How it works

The model is dense, which means it does not divide work between experts. It has training on 12 trillion tokens and handles a maximum of 1M tokens of context. It operates on local devices such as NVIDIA DGX Spark. It is on Hugging Face, Together AI, GMI Cloud and OpenRouter.

The numbers

The other scores are Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8 and NVIDIA Nemotron 3 Ultra 550B at 56.4. Ricursive says it used 1.4M training rollouts, about 70 times fewer than Reflection Beam.

What is not known

No external group did the tests. The company says the $32M cost is one tenth of the earlier best, “to the best of our knowledge”. This is a preview version.

Insight: Open model rivals now compete on cost, not size

The model of Ricursive is much smaller than the models it claims to beat. Vercel reports that open models are more than 7 in 10 of the tokens on its gateway, and Meta gave no test results for Muse Spark Open. This reads as a lab that uses a public model to prove its method, with chip design as its primary business. Test it on your own code tasks.