Ricursive releases 65B Ripple model for agentic coding

Ricursive Intelligence has released Ripple-65B Preview, a 65-billion-parameter dense open-weight language model trained from scratch for agentic coding. The model uses 12 trillion tokens, supports context lengths of up to 1 million tokens and is distributed under the Apache 2.0 license. The release includes instruction-tuned BF16, experimental NVFP4 and BF16 base checkpoints.

Ricursive reports a score of 80.52 on Terminal-Bench 2.1, the highest self-reported result among US open-weight models in its comparison. The company says Ripple exceeded the scores of competing mixture-of-experts models with 501B to 975B total parameters. Its other reported results include 79.1 on SWE-bench Pro v2.0 Hard, 76.5 on SWE-bench Verified and 98.3 on AIME 2025. Weaker results came on SciCode, at 39.4, and DeepSWE 1.1, at 44.3.

The model uses a dense Transformer with 80 layers and a 5120-wide hidden dimension. Ricursive says development cost $32M and used 1.4M reinforcement-learning rollouts. On two B200 GPUs, the BF16 model reaches 250 tokens per second at batch size eight. The NVFP4 version is intended for DGX Spark-class hardware, while weights are available through Hugging Face and listed launch providers.