Ricursive releases Ripple-65B Preview under Apache 2.0

Ricursive Intelligence released Ripple-65B Preview, a dense 65-billion-parameter open-weight model, under the Apache 2.0 license. The coding-focused release includes an instruction-tuned BF16 checkpoint that developers can download and run locally, subject to their hardware and serving setup.

The company also released an experimental NVFP4 instruction-tuned checkpoint and a BF16 base checkpoint. Ricursive says the NVFP4 checkpoint option applies four-bit NVFP4 precision to the model’s multilayer-perceptron layers while retaining BF16 attention. That lowers storage and memory requirements in part of the network without converting the entire model to four-bit precision. The base release is a pretraining-only checkpoint that excludes the midtraining and synthetic question-answer data used later in development.

The main model card describes Ripple-65B Preview as a thin-and-deep architecture with a model width of 5,120 and 80 layers. Ricursive says it pretrained the model from scratch on 12 trillion tokens, primarily English and code, and designed it for context lengths of up to 1 million tokens. The card does not establish a generally available serving configuration that exposes that full context window.

Ricursive reports scores of 80.52 on Terminal-Bench 2.1 and 79.1 on SWE-bench Pro v2.0 Hard. In the company’s model-card tables, Ripple ranks above four selected, larger U.S. open-weight models on both coding benchmarks. The comparisons have not been independently reproduced. The selected mixture-of-experts systems also have larger total parameter counts but activate fewer parameters per token than Ripple’s dense 65-billion-parameter design.

Ricursive also says its final Ripple v1 training ladder improved code compute efficiency by 16 times relative to its Olmo 3 ladder, with the advantage rising to 55 times at the full training run. The figures describe the company’s training methodology, not independently measured end-user coding performance, inference cost or operating cost.

The company cautions that modern agentic evaluations are difficult to reproduce because scores can shift with the evaluation harness, sequence length and summarization setup. It calls the benchmark table a best-effort fair comparison, not a settled independent ranking.