Ricursive: Ripple 65B Preview

Ripple-65B Preview is a 64.8B-parameter dense, Apache 2.0 model trained from scratch on 12 trillion tokens, with support for up to one million context tokens. Its checkpoints include an instruction-tuned BF16 model, an experimental NVFP4 version, a base model, and a 4B DFlash2 draft model intended to accelerate generation through speculative decoding. The review found that Ricursive's own comparison placed Ripple first on only two of 22 benchmarks, making its Terminal-Bench 2.1 advantage over Reflection Beam a narrow 0.4-point lead. The analysis says independent testing would be needed before treating the launch results as conclusive, and describes the model as particularly suitable for self-hosted coding workloads.

Ripple-65B Preview is a 65B dense model with 80 layers, pretrained on 12 trillion tokens and supporting up to one million context tokens. The article lists BF16 instruction, NVFP4, base, and DFlash2 checkpoints under Apache 2.0, and reports a creator-supplied Terminal-Bench 2.1 score of 80.5. A roughly 40 GB 4-bit GGUF could place the model on a single 48 GB graphics card or a 64 GB unified-memory system, broadening local agent use. The piece also reports more than 30 community-derived quantization repositories, but cautions that the benchmark comparisons were supplied by the lab and excluded some Chinese models.

Enjoy's analysis reports that Ripple-65B Preview's full weights occupy about 129.6 GB and shrink to roughly 37 GB in an unofficial 4-bit build, making the model practical on Macs with 64 GB or more memory. The model card describes a dense 65B architecture trained on 12 trillion tokens, with relatively light post-training focused on agentic coding. The review says Ricursive's 80.52 Terminal-Bench 2.1 score leads only within a narrowly defined US open-model comparison, and some competing results were measured by Ricursive itself. It also reports that the model was trained for $32 million and was released under Apache 2.0, though the supplied article excerpt is truncated.

Plotly evaluated Ripple-65B Preview on its 43-question data-analysis exam and found that it passed 35 questions, or 81%, at a reported cost of $2.21. It tied Kimi K3 and placed third among the tested US open-weight models, behind Muse Spark Open at 38 and Reflection AI's Beam 1.0 at 36. Ripple handled all five statistics traps and all five impossible requests, but found none of the five patterns hidden in the data. Plotly also highlighted its slower median completion time of 128 seconds per question and noted that each model was tested only once, limiting confidence in the ranking.

Ricursive Intelligence released Ripple-65B Preview, a 65-billion-parameter dense model pretrained on 12 trillion tokens with context support up to one million tokens. The creator says it is a general-purpose variant of the proprietary model powering its chip-design platform and reports 80.52 on Terminal-Bench 2.1, ahead of several larger open-weight models in its comparison. Ricursive says the model was developed and trained for $32 million and can run efficiently on local devices such as NVIDIA DGX Spark. The release is available through Hugging Face, Together AI, GMI Cloud, OpenRouter, and Fireworks. Its benchmark leadership and cost-efficiency claims are first-party statements rather than independently validated results.