Ornith-1.5-9B-GGUF by AtomicChat

Ornith-1.5-9B-GGUF by AtomicChat

GGUF builds of Ornith 1.5 9B, a model that reads images and text, in sizes from 2.81 GB to 17.9 GB plus a vision projector. Pick the file that fits your graphics memory and run it locally.

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: AtomicChat converted Ornith-1.5-9B, a 9B model that takes images and text, into GGUF files for local runners. The files range from 2.81 GB to 17.9 GB. The AD files use a bit layout for each tensor, tuned for this model and built with the team's own importance matrix. A separate vision projector file adds image input. The Atomic Chat app offers toggles for thinking.

Who it suits: It suits people who want to run a 9B image and text model on their own graphics card and need a file that matches their memory.

Industry | Who uses it | How they use it | Example
Local AI | People running models offline | Choose the file that fits their graphics memory | Take the 5.93 GB file on an 8 GB card
Vision tasks | Users who ask about pictures | Load the vision projector beside the language file | Ask the model about a screenshot
Research | Quantization researchers | Compare AD files with stock presets that share one matrix | Read the KL divergence and top-1 columns

You can skip it if you have less than 4 GB of video memory, since the card's smallest suggestion is for 4 GB.

Where to use it:

• llama.cpp, the runner the card measured with
• Atomic Chat, the app from the same team
• Windows, macOS and Linux with Python, to download the files

Use case | Who | What they get back
Local assistant | Developers and hobbyists | A 9B model that answers in text on their own machine
Picture questions | Users with the projector file | Answers that take an image into account
File choice | People with limited memory | A file matched to 4, 6, 8, 12, 16 or 24 GB

How to set it up: You need these first:

• Python, to use Hugging Face's download client
• Graphics memory that fits the file you pick
• llama.cpp or the Atomic Chat app

1. Install or update the Hugging Face download client.
2. Download the repository's files to your computer.
3. Pick a file from the card's memory table.
4. Load it in your runner, and add the vision projector if you want images.

pip install -U huggingface_hub
hf download AtomicChat/Ornith-1.5-9B-GGUF

How to use it: You load one language file in llama.cpp or Atomic Chat and talk to the model. Leave out the projector for text only. The text available here cuts off before the run steps, so follow the model card and its guide for the exact options.

Tool | What it does
BF16 file | The reference build at 17.9 GB, for 24 GB and up
Q8_0 file | A 9.53 GB build for 16 GB
AD-Q5_K-Q4_K file | A 5.93 GB build for 8 GB, which the card calls the sweet spot
AD-Q4_K-IQ4_XS file | A 5.61 GB build for a tight 8 GB
Vision projector file | A 0.92 GB file that adds image input

Example: You have an 8 GB graphics card, so you take the 5.93 GB AD-Q5_K-Q4_K file and load it with the vision projector. The card measured this file at a top-1 match of 93.10% against the reference.

For newcomers and experts: A smaller file fits less memory but loses some accuracy. A good first try is the file the card lists for your amount of video memory, with a short chat.

For readers who know the field, the card reports two measures against the BF16 reference. Mean KL divergence is better when lower. Top-1 counts how often a file chooses the same next word as the reference does. Tests used a held-out neutral text set at 4096 context on one RTX 5090 with llama.cpp b10505. The plain presets were built with that same matrix, so any gap comes from the AD layout alone.

Check your graphics memory
Pick a file from the card's table
Download it
Load it, with the projector for images
Chat with the model

Good to know:

• Sizes: the listed sizes already count the 0.9 GB projector and about 1 GB of overhead at 8k context.
• Smallest file: the card publishes its two smallest files for completeness and says the smallest acts like a different model.
• Warning: the 2.81 GB file agrees with the reference on only 53.74% of tokens.
• Cut: the text available here ends before the card's chart discussion.
• Counts: Hugging Face lists 9B parameters and about 4.5 GB at 4-bit as an estimate.

Title: Top-1 agreement with the BF16 reference, higher is better
Q8_0: 97.94%
Q6_K: 96.54%
AD-Q5_K-Q4_K: 93.10%
Q5_K_M: 92.80%
AD-Q4_K-IQ4_XS: 91.93%
IQ3_M: 84.18%
AD-IQ2_XXS-IQ1_M: 53.74%

• License: MIT.
• Cost to run: none named. You need graphics memory that fits the file.
• Data it reaches: the prompts and images you give it on your own machine. The text available here names no network use.
• Not for: real work with the two smallest files, which the card publishes only for completeness.

On Hugging Face, 14 November 2026

• Task: image text to text
• Library: gguf
• Parameters: 9B
• Weights: about 4.5 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 41,746
• Likes: 83
• License: mit
• Based on: ornith-ai/Ornith-1.5-9B
• Tags: gguf, llama.cpp, vision, agentic-coding, imatrix, conversational, image-text-to-text

Links

• Model: https://huggingface.co/AtomicChat/Ornith-1.5-9B-GGUF
• Files: https://huggingface.co/AtomicChat/Ornith-1.5-9B-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download AtomicChat/Ornith-1.5-9B-GGUF

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