Ripple-65B Preview NVFP4 (Experimental)

Ricursive Logo

BF16 | NVFP4 | Base Model | DFlash2 | Launch Blog | License: Apache 2.0

Ricursive Intelligence introduces Ripple-65B Preview, a 65B parameter dense Transformer model pretrained on 12 trillion tokens with support for up to 1M context length. 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), Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8), and NVIDIA Nemotron 3 Ultra 550B (56.4).

Our Preview release includes an instruction-tuned model in BF16 and NVFP4 numerics, and a BF16 base model. This repo contains the NVFP4 instruction-tuned checkpoint.

This is a conservative quantization of the BF16 checkpoint, using NVFP4 for the MLP layers while keeping attention in BF16.

Run on DGX Spark

inference/local/ serves this checkpoint with the DFlash2 draft on a single DGX Spark (GB10, 128 GB unified memory) through stock, unmodified SGLang 0.5.19. You need Python 3.12, the CUDA 13 toolkit, build-essential and python3.12-dev.

hf download ricursiveAI/Ripple-65B-Preview-NVFP4 --include "inference/local/*" --include chat_template.jinja --local-dir ripple
cd ripple/inference/local
python3.12 -m venv .venv && .venv/bin/pip install -r requirements.txt
CUDA_HOME=/usr/local/cuda-13.0 PATH="$PWD/.venv/bin:/usr/local/cuda-13.0/bin:$PATH" python spark.py

spark.py downloads the checkpoints and starts SGLang's OpenAI-compatible server at http://127.0.0.1:8000/v1 with model name ripple. It supports streaming, tool calls and reasoning_content. ripple.ipynb uses the same engine from a notebook. In Python scripts, create Spark(...) under if __name__ == "__main__":, because SGLang starts worker processes.

curl http://127.0.0.1:8000/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model":"ripple","messages":[{"role":"user","content":"Write a Python binary search."}],"max_tokens":256}'

Defaults:

  • Decoding: speculative tree decoding with 32 verification nodes from a 16-token draft, using the bundled kernels and deterministic Triton attention. Tree decoding is experimental. --decoding chain --kernels stock selects stock SGLang chain speculation.
  • Context: 131,072 tokens (input up to 131,065), with prefix caching on, so follow-up turns reuse their cached prefix.
  • Concurrency: up to 4 concurrent requests.
  • Sampling: temperature 1.0 and top-p 0.95 when a request omits them. Send "temperature": 0 for greedy output.

On a DGX Spark at temperature 1.0, this reaches about 19.5 tok/s for one request and 48 tok/s in total with four, on representative prompts with 128-token outputs.

Ripple-65B-Preview: NVFP4 vs BF16 benchmarks

Benchmark Ripple-65B
NVFP4
Ripple-65B
BF16
Nemotron 3
Ultra 550B BF16
Inkling
975B BF16
Muse Spark
Open 816B BF16
Reflection
Beam 501B BF16
Model
Architecture Dense Dense MoE MoE MoE MoE
Total params 65B 65B 550B 975B 816B 501B
Active params 65B 65B 55B 41B 42B 23B
Pre-training tokens 12T 12T 20T 45T N/A 23.8T
Knowledge
MMLU 89.2 89.6 92.0 91.3 92.6 92.4
MMLU-Pro 83.7 84.2 86.8 87.3 87.9 76.5
GPQA Diamond 83.3 85.2 87.0 87.2 89.0 90.6
Math
AIME 2024 97.1 97.1 94.2 97.5 97.9 96.7
AIME 2025 97.9 98.3 94.2 91.7 100.0 99.6
AIME 2026 95.8 95.0 95.4 97.1 98.3 97.7
HMMT Feb 2026 89.4 90.5 80.3 86.3 93.9 93.9
Coding & Agentic
Terminal-Bench 2.1 (self-reported) 76.0 80.5 56.4 63.8 76.4 80.1
Terminal-Bench 2.1 (Terminus 2) 76.0 80.5 55.1 56.2 78.7 70.8
SWE-bench Verified 76.5 76.5 71.9 77.6 78.7 81.6
SWE-bench Pro v1.0 61.06 61.42 46.9 54.3 58.2 66.1
SWE-bench Pro v2.0 Hard 77.8 79.1 30.7 39.6 72.4 76.5
DeepSWE 1.1 40.3 44.3 0.2 1.8 60.0 43.3
SciCode (Inspect AI) 39.5 39.4 44.6 46.1 58.5 49.2
ฯ„ยณ-bench banking 29.90 31.27 22.6 23.7 14.4 36.1
AutomationBench (Public) 36.3 36.4 8.4 14.6 42.4 39.3
MCP Atlas 73.27 71.67 62.2 76.0 89.2 80.8
Safety
FORTRESS Adversarial 83.2 83.8 79.3 78.0 94.2 92.1
FORTRESS Benign 95.5 95.5 89.1 95.9 82.6 86.4
StrongREJECT 95.9 97.5 98.2 98.6 94.7 95.8
Instruction Following
IFBench 78.6 79.2 81.7 79.8 78.6 79.6
Long Context
AA-LCR 77.3 75.0 65.4 78.3 80.7 78.2

Citation

@misc{ricursive2026ripple65b,
      title={Ripple-65B: Pushing the Frontier of Local Intelligence},
      author={Ricursive Intelligence},
      year={2026},
}

Contact

For questions and comments, please contact us at info@ricursive.com. To learn more, please visit our website.

Downloads last month
5,913
Safetensors
Model size
40B params
Tensor type
BF16
ยท
U8
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ricursiveAI/Ripple-65B-Preview-NVFP4

Finetunes
1 model