Ricursive Intelligence, an AI chip design firm, trained Ripple-65B Preview on 12 trillion tokens. The company says the model reached 80.52 on TerminalBench 2.1. Ripple uses Terminus 2 and OpenCode as its harness, partly relying on each.
Ripple does not match the newly released Inkling 2 Small on the other benchmarks reported. On SWE-Bench Pro v2.0 Hard it scores 79.1% against 86.3%, and on DeepSWE v1.1 it scores 44.3% against 72.1%. The model is a dense Transformer with 65 billion parameters rather than a mixture-of-experts design, and it is optimized for Nvidia hardware.
Ricursive says total development and training costs came to $32 million. The weights are available on Hugging Face under the Apache 2.0 license. The Decoder notes that Ripple is likely easier to deploy on users' own hardware than Inkling 2 Small.
Why it matters if you build agents
Developers who want a self-hosted coding model now have an Apache 2.0 option with strong reported agentic coding results. The scores are company-reported and the model trails Inkling 2 Small on the benchmarks shown.
Test Ripple-65B Preview in your own harness on your own tasks before choosing it over a larger or closed model.
Get stories like this every weekday
One short email with the day's agent tooling news.