Ripple-65B model card claims an 80.5 to 80.1 Terminal-Bench lead and fills rival gaps with in-house runs

Ricursive Intelligence released Ripple-65B Preview on Hugging Face on Nov 4 under Apache 2.0. It is a 65B-parameter dense Transformer pretrained from scratch on 12 trillion tokens, with up to 1M tokens of context. The release includes BF16 and NVFP4 instruct checkpoints and a BF16 base model. Ricursive says the full R&D cost was $32M and that it reached 80.5 on Terminal-Bench 2.1 with 1.4M RL rollouts. The model card compares Ripple with four U.S. open-weight MoE models of 501B to 975B total parameters, each with 23B to 55B active parameters against Ripple's 65B. Ricursive says the table uses official model-card scores where they exist and its own eval runs, set in italics, for the missing ones. On Terminal-Bench 2.1, Ripple's 80.52 edges Reflection Beam 501B's self-reported 80.1. In Ricursive's runs with the official Terminus 2 harness, Beam scores 70.8 and Ripple 80.5. Ricursive says most models scored lower under that harness. For SWE-bench Pro v2.0 Hard and DeepSWE 1.1, Ripple's own scores also come from Terminus 2. The weights are on Hugging Face, and Ricursive says the model is also served on Together AI, GMI Cloud and OpenRouter. Why it matters: A 65B dense model that keeps up with MoE models of 8 to 15 times its total size on coding-agent tests would let more teams run a capable coding agent on hardware they control. Ricursive says the model can fit on an NVIDIA DGX Spark, and it ships an NVFP4 checkpoint for local use. The headline lead is 0.4 points, 80.5 to 80.1, and both numbers are the labs' own scores for their own models. Ricursive itself writes that "replicating benchmark scores has been challenging" for agentic evals. Several of the weakest rival scores, such as 0.2 for Nemotron 3 Ultra on DeepSWE 1.1 and 2.0 for Inkling on SWE-bench Pro v2.0 Hard, come from Ricursive's in-house runs. Until someone else reruns them, treat the italic cells as Ricursive's measurements, not the rivals' results. The weights are public, so you can test the model on your own tasks.

Try this: Before you switch models, run Ripple-65B Preview through OpenRouter on a few real tasks from your own repo, inside the coding agent you already use, next to your current model. Compare pass rates and run time. When you quote the model card, check whether the cell is in italics, which marks Ricursive's own eval run.