Swift 1.5 Ornith 9B GGUF

Swift 1.5 Ornith 9B GGUF

Runs Swift 1.5 Ornith 9B, a reasoning model tuned to think with fewer tokens, as GGUF quantizations for llama.cpp. For people who run a 9B reasoning model on their own machine.

Made by ukisai 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 quantizations of Swift 1.5 Ornith 9B, a 9.2B model from UkisAI. Swift is a reasoning-efficient version of Ornith-1.5-9B. The card reports 33.1 percent fewer thinking tokens than that base, along with a GPQA-Diamond score 1.52 points higher. The files were made directly from the Safetensors release with llama.cpp.

Who it suits: It suits developers and hobbyists who want a 9B reasoning model on local hardware, and anyone for whom long thinking traces cost too much time.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a local reasoning model through llama-server | Answer a hard question with a shorter thinking trace
Research | ML researchers | Compare a token-penalty fine-tune with its base | Set Swift 1.5 next to Ornith-1.5-9B
Self-hosting | Hobbyists | Load one quantization tier on a modest machine | Run a 4-bit class file in a llama.cpp runtime

People who want the original Safetensors weights can skip this repository, since the card points to a separate one for those.

Where to use it:

• llama.cpp and runtimes that are compatible with it, such as llama-server
• Windows, macOS or Linux, where the files download with Python

Use case | Who | What they get back
Answer reasoning questions locally | Developers | A reply that spends fewer thinking tokens than the base
Test a quantization tier | Researchers | A file you can load and judge on your own tasks
Offline question answering | Hobbyists | A model on your machine with no hosted service

How to set it up:

• Python with pip, for the download
• A current runtime that is compatible with llama.cpp, such as llama-server
• Disk space for the tier you pick, about 4.6 GB at 4-bit as an estimate

1. Install the Hugging Face hub client.
2. Download the repository's files.
3. Start your runtime on the quantization tier you chose.

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

How to use it: Choose one of the quantization tiers, then start it in llama-server or another recent runtime that reads llama.cpp files. After it loads, you send prompts the way you would to any local chat model.

Tool | What it does
hf download | Fetches the repository's files to your computer
llama-server | A llama.cpp runtime that can load a GGUF tier and answer prompts

Example: You load a tier in llama-server and ask a hard science question of the kind GPQA-Diamond uses. The card says the model produces shorter reasoning traces than its base, so the answer should arrive after fewer thinking tokens.

For newcomers and experts: A reasoning model writes out its thinking before it answers, and some of them think for too long. This one was tuned to think less. Download the files, load one tier in llama-server and ask it a question you can check.

The makers found tokens linked to overthinking and penalized them, instead of attacking reasoning length directly, which gave compressed token use. Their training data was a multi-turn agent set that they re-sampled and turned into RL environments. They also report fewer overthinking errors in their own testing.

Start from Ornith-1.5-9B
Penalize tokens tied to overthinking
Get Swift 1.5 Ornith 9B
Convert to GGUF tiers with llama.cpp
Run a chosen tier locally

Good to know:

• Reported gains: 33.1 percent fewer thinking tokens and a 1.52 point higher GPQA-Diamond score than the base, as the card states
• Size: 9.2B parameters, with about 4.6 GB at 4-bit as an estimate
• Tag: the page lists it as image text to text, but the card does not show image use
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the prompts you send to your own runtime, with no web or account access named
• Not for: nothing named in the card

On Hugging Face, 14 November 2026

• Task: image text to text
• Library: gguf
• Parameters: 9.2B
• Weights: about 4.6 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 3,763
• Likes: 26
• License: apache-2.0
• Based on: ukisai/Swift-1.5-Ornith-9B
• Tags: gguf, llama.cpp, swift, ornith, qwen3_5, reasoning, efficient-thinking, token-efficient, post-training, experimental, image-text-to-text, conversational

Links

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

Get the model

Download its files with Hugging Face's command line:

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

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