Ornith-1.0-9B-GGUF

Ornith-1.0-9B-GGUF

GGUF copy of a 9B open-source coding-agent model, trained with reinforcement learning to write solutions and the scaffolding around them. Built for single-GPU use by developers running agents locally.

Made by ornith-ai 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: Ornith-1.0-9B is the smallest model in the Ornith family, a set of open-source coding-agent models that the card calls self-improving. This repository holds its GGUF copy, which the card says is designed for efficient use on a single GPU. The family's base models are Gemma 4 and Qwen 3.5, which the authors trained further.

Who it suits: It suits developers who want a coding agent model on their own machine and researchers who follow how reinforcement learning is applied to agents.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a local coding agent on one GPU | Ask it to fix a failing test in a repository
Developer tools | Tool builders | Connect the model to an agent setup of their own | Test it in a terminal-based coding agent
Research | ML researchers | Study training that improves both the scaffold and the solution | Compare it with the Qwen3.5-9B scores in the card

Anyone who needs a chat model for writing or general questions can skip it, since the card positions it for agentic coding.

Where to use it: You run it on your own computer, behind an agent setup that calls the model.

• Windows, macOS or Linux, with Python, for downloading the files
• A single GPU, the deployment the card designs it for
• Terminus-2 and Claude Code, the two agent setups named in the card's benchmark table

Use case | Who | What they get back
Fix a bug in a repository | Developers | Edits and commands proposed across several steps
Run terminal tasks | Agent builders | A model to test on terminal-based task suites
Study agent training | Researchers | A small checkpoint trained with reinforcement learning

How to set it up: You need a way to fetch the files, a GPU and a runtime that opens GGUF files.

• Python, for the Hugging Face hub tools
• A single GPU
• Free disk space, about 4.5 GB for a 4-bit copy by estimate
• A local runtime that reads GGUF files

1. Install the Hugging Face hub tools.
2. Download the repository files.
3. Open the file in your GGUF runtime and connect it to your coding agent setup.

pip install -U huggingface_hub
hf download ornith-ai/Ornith-1.0-9B-GGUF

How to use it: The part of the card shown here gives no run command, so use is a matter of loading the file in your runtime and pointing your agent at it. Give the agent a coding task and let it work through the steps.

Example: You hand your agent the task, "Make the failing test in this repository pass." The model reads files, runs commands and proposes edits over several turns, and your agent applies them.

For newcomers and experts: A coding agent does more than write code. It runs commands and edits files step by step through an app, and this model is built to drive one. A good first try is to load the file and ask it a short coding question before wiring it into an agent.

For someone who knows the field, the card says the training uses reinforcement learning on two things at once: the solution attempts and the scaffold that drives them. Optimizing both together is meant to find better search paths. The card reports results on four coding benchmarks, among them SWE-Bench and OpenClaw, and those figures are the authors' own.

The model proposes a scaffold
The scaffold drives solution attempts
Reinforcement learning scores the pair
Training favors better search paths

Good to know: The card's table gives 69.4 on SWE-bench Verified for this model against 53.2 for Qwen3.5-9B. The chart shows the Terminal-Bench 2.1 scores with the Terminus-2 setup, as the card reports them.

Title: Terminal-Bench 2.1 score with Terminus-2
Ornith-1.0-9B: 43.1 points
Qwen3.5-9B: 21.3 points
Qwen3.5-35B: 41.4 points
Gemma4-12B: 21 points
Gemma4-31B: 42.1 points

• License: MIT, which the card describes as open worldwide with no regional limits
• Cost to run: none named, but you need a GPU
• Data it reaches: it downloads files from Hugging Face. What the model touches while coding depends on the agent setup you connect.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 9B
• Weights: about 4.5 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 2,357,551
• Likes: 702
• License: mit
• Quantized versions: 2
• Tags: transformers, gguf, text-generation, conversational

Links

• Model: https://huggingface.co/ornith-ai/Ornith-1.0-9B-GGUF
• Files: https://huggingface.co/ornith-ai/Ornith-1.0-9B-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ornith-ai/Ornith-1.0-9B-GGUF

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