Ricursive's Ripple-65B tops US open-weight models on coding
A 65B dense model trained for $32 million scores 80.52 on TerminalBench 2.1, beating open-weight rivals up to 816 billion parameters, and runs locally on an Nvidia DGX Spark.
Ricursive Intelligence released Ripple-65B Preview on 4 November, a 65-billion-parameter open-weight model aimed squarely at agentic coding. It’s a dense Transformer, not the mixture-of-experts design most labs now favour, with 80 layers and a d_model of 5120, pretrained on 12 trillion tokens with context extending to 1M.
The model scores 80.52 on TerminalBench 2.1, which Ricursive says is state of the art among US open-weight models, ahead of Reflection’s 501B-parameter Beam (80.1), Meta’s 816B Muse Spark Open (76.4) and Nvidia’s 550B Nemotron 3 Ultra (56.4). Ricursive says the full build, from pretraining ladder development through the final run, cost $32 million, about a tenth of what it estimates the previous open-weight leader spent. The small parameter count means it can run locally on an Nvidia DGX Spark. Weights are live now on Hugging Face, Together AI, GMI Cloud and OpenRouter, with Fireworks following later the same day.
Why it matters: a 65B dense model beating open-weight rivals eight to twelve times its size, for a fraction of the compute spend, is a real data point against the assumption that frontier coding performance needs a frontier-sized budget. Worth watching if you’re choosing a model to self-host for agent work.
The caveat: every figure here, the benchmark score, the cost, the comparisons, is Ricursive’s own. There’s no independent benchmark run yet, and the page doesn’t name a specific open-weight licence.