How to run Ripple-65B Preview: OpenRouter, Hugging Face, DGX Spark

You can run Ripple-65B Preview three ways: call it through OpenRouter, download the weights from Hugging Face, or load the smaller NVFP4 checkpoint on a machine like the NVIDIA DGX Spark. The model is the 65B open-weight coding model that Ricursive Intelligence released on November 4, and our news article covers what it is. This guide covers how to use it.

None of the commands below were run for this article. They come from the OpenRouter and Hugging Face pages as read on November 5, 2026, so check the pages if something fails.

Option 1: Call Ripple-65B Preview through OpenRouter

This is the quickest route, with no hardware needed. OpenRouter's model page lists the model id ricursiveai/ripple-65b-preview, served by GMI Cloud, through the standard chat completions endpoint.

1. Create an OpenRouter account and an API key.
2. Send a request to https://openrouter.ai/api/v1/chat/completions with the model id above.

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ricursiveai/ripple-65b-preview",
    "messages": [{"role": "user", "content": "Write a Python function that parses ISO dates."}]
  }'

Streaming works, and the page lists support for tools, tool_choice, response_format, seed, temperature, top_p and logprobs. As of November 5, 2026, the price is $0.20 per million input tokens and $0.60 per million output tokens, with no markup from OpenRouter. GMI Cloud is the only provider named, and EU in-region routing is not available for this model.

Option 2: Download the weights from Hugging Face

The model card says the weights are BF16 Safetensors under the Apache 2.0 license, so you can use them commercially. The repository also has an NVFP4 checkpoint, a base model and a DFlash2 variant. Two cautions from the card:

• The repository is tagged custom_code, so loading it may require trusting remote code in your inference library.
• No Hugging Face inference provider currently hosts it.

Option 3: Run it locally

Ricursive says the model is small enough to run on an NVIDIA DGX Spark, and the card points to an NVFP4 checkpoint and a Colab notebook for getting started. The card gives no minimum hardware requirements. Its speed figure, 250 tokens per second, is for two B200 GPUs at batch size 8, which is far more than a desktop machine.

Which option to pick

Limits to know before you rely on it

This is a preview with light post-training, aimed at agentic coding, and its training data is mostly English. Its 80.5 on Terminal-Bench 2.1 is the lab's own number, and the authors say scores are lower under the official Terminus 2 harness at 128K context.