Ricursive says Ripple-65B Preview scores 80.52 on Terminal-Bench 2.1, ahead of much larger US open-weight models such as Reflection Beam 501B (80.1) and Meta Muse Spark Open 816B (76.4). That's the lab's own figure from its model card.
The release has an instruction-tuned model in BF16 (about 130 GB on the Hub), an experimental NVFP4 build (about 57 GB) that Ricursive says fits on an NVIDIA DGX Spark, and a BF16 base model. The card quotes 250 tokens per second on two B200 GPUs at batch 8, before quantization.
One catch: it's a custom architecture. Ricursive's own server code plugs it into SGLang 0.5.19, and Transformers needs trust_remote_code. If you want a hosted copy, ByteFwd reported OpenRouter listing it on Nov 5 at $0.20 per million input tokens and $0.60 per million output tokens.