Ripple-65B: a dense open model taking on big MoEs

What Ricursive released

Ricursive Intelligence published Ripple-65B Preview on Hugging Face under Apache 2.0. It is a 65-billion-parameter dense Transformer with 1,048,576 tokens of context. The release includes an instruction-tuned model in BF16 and NVFP4, plus a BF16 base model.

It was pretrained from scratch on 12 trillion tokens of mostly English text and code. The team grew from one to five people during development.

Why dense instead of MoE

Most large open-weight models today use Mixture-of-Experts. Ricursive chose a dense model to keep training infrastructure simple: plain sharding, no custom kernels and no quantization, all in BF16. It says that kept MFU at about 40% through pretraining and post-training.

The architecture is thin and deep, with a 5120 d_model and 80 layers. The team started from Olmo 3, then moved to a Gemma 3N-like design for stability, adding gated attention, global NoPE, value residuals and FFN activation clipping. Ricursive reports a 16x compute-efficiency gain on code over the Olmo 3 ladder, rising to 55x for the full run.

What the numbers show

Ripple-65B scored 80.5 on Terminal-Bench 2.1. The model card says that beats the self-reported scores of Reflection Beam 501B (80.1), Muse Spark Open 816B (76.4), Inkling 975B (63.8) and Nemotron 3 Ultra 550B (56.4).

Other figures: SWE-bench Verified 76.5, SWE-bench Pro v2.0 Hard 79.1, GPQA Diamond 85.2, AIME 2026 95.0 and AA-LCR 75.0. It does not win everywhere. On SciCode it scored 39.4 against 58.5 for Muse Spark Open, and on MCP Atlas 71.7 against 89.2.

Every number is self-reported, and missing values come from Ricursive's in-house eval flow. The company itself notes that agentic evals are hard to replicate because many factors shift scores. There is no paper or independent evaluation yet.

Where it runs

At 65B it fits on an NVIDIA DGX Spark, and Ricursive ships an NVFP4 checkpoint and a Colab notebook for that. On 2xB200 at batch size 8, without quantization, it measured 250 tokens per second. A community GGUF build for llama.cpp is also available.

Why it matters

If you want a local agentic coding model, Ripple-65B is an interesting alternative to MoEs with hundreds of billions of parameters: a single dense model is simpler to serve, with no expert routing. Muse Spark Open, which we covered earlier, needs two Mac Studios: Meta's Muse Spark Open: open weights that run on Macs.

This is a preview, and Ricursive says post-training is still light. Test it on your own tasks before swapping anything out.