Ricursive opens Ripple-65B after training it for $32 million
The 65B dense model uses an Apache 2.0 license and is built for local agentic coding, though its benchmark lead is based on vendor-reported results.
Ricursive Intelligence has released Ripple-65B Preview, a 65-billion-parameter dense language model aimed at reasoning and agentic coding. The weights are available under Apache 2.0, including BF16, base-model and NVIDIA NVFP4 variants.
The company says it pretrained Ripple from scratch on 12 trillion mostly English and code tokens, then added long-context and reinforcement-learning stages. The model supports contexts of up to one million tokens. At 65 billion parameters it is still large, but Ricursive says the quantized build can fit on hardware such as NVIDIA’s DGX Spark, making it substantially easier to self-host than several competing models with hundreds of billions of total parameters.
Ricursive reports a score of 80.52 on TerminalBench 2.1, compared with 80.1 for Reflection’s 501B-parameter Beam model. The company describes this as the best result among US open-weight models. The difference is small, and the comparison mixes official model-card numbers with Ricursive’s own evaluations, so it should be treated as a vendor claim rather than an independent ranking.
The company also says Ripple’s full research and development cost was $32 million, about one-tenth of the previous open-weight leader by its estimate. That claim includes more than compute: Ricursive says the model was built by a team that grew from one to five people.
The practical value is the combination of a permissive license, local deployment and strong coding results. Ripple is also offered through Together AI, GMI Cloud and OpenRouter for teams that do not want to host it themselves.
