Ripple-65B-Preview-GGUF

Ripple-65B-Preview-GGUF

GGUF quantized files of the 65B Ripple Preview agentic-coding model, for running locally with llama.cpp. For people who run large open models on their own hardware.

Made by AtomicChat on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository holds GGUF files of Ripple-65B Preview, a 65B dense model that Ricursive Intelligence trained from scratch on 12T tokens and post-trained for agentic coding. Atomic Chat made the files from the original BF16 weights and wrote the llama.cpp support for it. The set runs from a bit-exact BF16 copy down to small quantized files.

Who it suits: It suits developers and hobbyists with a large GPU box or a high-memory Mac who want a coding model that runs locally.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a coding model on their own machine | Pick a file that fits a 64 GB Mac
AI research | Quantization researchers | Compare file layouts by KL divergence | Read the KLD and same-top-token columns side by side
Local AI | Hobbyists with big GPUs | Match a file to their memory | Take AD-IQ3_S on a 36-48 GB setup
Privacy-minded teams | Engineering leads | Keep code and prompts on their own hardware | Run the model offline in a local app

If your machine has less than about 24 GB of memory, you can skip it.

Where to use it:

• llama.cpp build b12167 or newer, as a release build or from source
• LM Studio and Ollama, once their bundled llama.cpp reaches b12167 or newer
• Atomic Chat, an app from the same team for running open-weight models on your own machine without internet
• Windows, macOS or Linux, with enough memory for the file you pick

Use case | Who | What they get back
Local agentic coding | Developers | A 65B coding model answering on your own hardware
Choosing a quantization | Local AI users | A file matched to your memory, from 18.5 GB to 129.6 GB
Quality comparison | Researchers | Measured KLD and token agreement against the BF16 original

How to set it up: You need these first:

• llama.cpp b12167 or newer, with no patch
• Python, for the Hugging Face download command
• Memory and disk that fit the file you pick, as the chart below shows

1. Pick a file from the table, since the repository holds all of them and the sizes differ a lot.
2. Install the Hugging Face hub tool and download the files.
3. If a file comes in several parts, give llama.cpp the first part.

pip install -U huggingface_hub
hf download AtomicChat/Ripple-65B-Preview-GGUF

How to use it: Load the file you chose in llama.cpp, or in an app that bundles a recent enough llama.cpp. The quality figures were measured with Ripple's own chat template, so use that template when you chat.

Title: Some of the files, with their size and fit
File | GB on disk | Suggested hardware | Same top token
BF16 | 129.6 | The original weights, in 3 parts | 100%
Q8_0 | 68.8 | 128 GB Macs and DGX Spark | 99.0%
AD-Q4_K_M | 40.4 | 48 GB, such as two 24 GB GPUs | 95.9%
AD-IQ3_S | 32.4 | 36-48 GB | 93.4%
AD-IQ2_XXS | 20.0 | 24 GB, such as an RTX 3090 or 4090 | 80.8%
AD-IQ1_M | 18.5 | Research use only | 76.8%

Example: You have a 64 GB Mac and pick AD-Q5_K_M, which takes 47.9 GB on disk. You load it in llama.cpp and ask the model to review a function. Per the table, it picks the same top token as the BF16 original 97.3% of the time.

For newcomers and experts: GGUF is a file format for running models on your own computer, and this repo offers the same model at many sizes. A good first move is to choose the biggest file your memory allows, then try it in a llama.cpp chat.

The AD files set a layout for each tensor. They use an importance matrix worked out from the BF16 weights on a calibration set of 4.7M tokens. Quality is scored by mean KL divergence and by how often a file picks the BF16 model's top token. At the size of plain Q4_K_M, AD-IQ4_XS has 2.5x lower mean KLD, at 0.0149 against 0.0380.

Pick a file that fits your memory
Download it with the hub tool
Load it in llama.cpp b12167 or newer
Chat using Ripple's chat template

Good to know: File size grows quickly with precision, so the biggest files need server-class memory.

Title: File size on disk in GB
BF16: 129.6 GB
Q8_0: 68.8 GB
AD-Q6_K: 57.7 GB
AD-Q5_K_M: 47.9 GB
AD-Q4_K_M: 40.4 GB
AD-IQ3_S: 32.4 GB
AD-IQ2_XXS: 20.0 GB
AD-IQ1_M: 18.5 GB

• Context: the hardware suggestions assume a medium-length context, and a longer one needs extra memory for the KV cache
• Speed: in one run, the plain Q4_K_M on one GH200 96 GB processed the prompt at 309 tokens/s and generated 40.6 tokens/s
• Research file: AD-IQ1_M is meant for research, as it matches the original's token choice only 76.8% of the time
• License: Apache 2.0
• Cost to run: none named, you supply the hardware
• Data it reaches: the model files you load, since the model runs locally

On Hugging Face, 13 November 2026

• Task: text generation
• Library: gguf
• Parameters: 64.8B
• Weights: about 32.4 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 100,818
• Likes: 53
• License: apache-2.0
• Based on: ricursiveAI/Ripple-65B-Preview
• Tags: gguf, llama.cpp, ripple, ricursiveai, quantized, conversational, agentic-coding, atomic-chat, text-generation, imatrix

Links

• Model: https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF
• Files: https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download AtomicChat/Ripple-65B-Preview-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.