Ricursive's Ripple-65B tops bigger US open-weight models

Ricursive Intelligence has released Ripple-65B Preview on Hugging Face, a 65-billion-parameter dense Transformer under an Apache 2.0 license. The model card reports a score of 80.52 on Terminal-Bench 2.1.

"Ripple-65B Preview delivers strong results on agentic coding, scoring 80.52 on TerminalBench 2.1 and surpassing significantly-larger US open-weight models such as Reflection Beam 501B (80.1), Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8), and NVIDIA Nemotron 3 Ultra 550B (56.4)," the card states.

The architecture is dense rather than Mixture-of-Experts: a 5120 d_model, 80 layers, and context length up to 1,048,576 tokens, a design the card describes as Gemma 3N-like. Pretraining ran on 12 trillion tokens of English and code. Other reported benchmarks include SWE-bench Verified at 76.5, MMLU at 89.6, GPQA Diamond at 85.2, AIME 2026 at 95.0, and AA-LCR at 75.0 on long context.

The efficiency pitch is central. The card claims 40% MFU through pretraining and post-training, 250 tok/s on 2xB200 at batch size 8 prior to quantization or custom kernels, and compute-efficiency improvements of 16x on code over an Olmo 3 ladder, rising to 55x for the full hero run with scaling. Ricursive says the team scaled from one to five people during development, and chose dense over MoE for "no custom kernels, no quantization required (BF16)," with the model sized to fit on an NVIDIA DGX Spark.

Two researchers we track shared the release link. Every benchmark here is self-reported; the preview ships without an accompanying paper or external evaluation.