Ornith-1.5-35B-A3B-NVFP4

Ornith-1.5-35B-A3B-NVFP4

Writes code and handles terminal agent tasks, a mixture-of-experts model using about 3B parameters per token, packaged as an NVFP4 copy. For developers hosting a coding model on their own hardware.

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: This repository holds Ornith-1.5-35B-A3B from ornith-ai, and its name points to an NVFP4 quantized copy. The model is aimed at coding and agent work. It is a mid-size mixture-of-experts model that uses about 3B parameters for each token, and the files weigh about 19.5 GB.

Who it suits: It suits developers and teams who want a coding and agent model they can host themselves, and researchers who follow how models are trained.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a coding model on their own hardware | Asking it to fix a failing task in a terminal agent
Research | Machine learning researchers | Study a model trained with a self-improvement loop and reinforcement learning | Comparing it with Ornith-1.0 on terminal tasks
Tooling | Agent builders | Use it as the model behind terminal based agents | Running it under Terminus-2 or Claude Code, the two setups the card scores

You can skip it if your work is not coding or agent tasks, or if your machine cannot hold about 19.5 GB of weights.

Where to use it:

• Python with the transformers library, on Windows, macOS or Linux
• Terminal agent setups, such as the Terminus-2 and Claude Code runs the card reports
• Your own computer, with memory for the weights

Use case | Who | What they get back
Coding help | Developers | Code and fixes written by the model
Terminal tasks | Agent builders | Command line steps from an agent run
Model comparison | Researchers | Benchmark rows to set against Ornith-1.0 and Qwen3.6-35B-A3B

How to set it up:

• Python and pip
• Disk space and memory for about 19.5 GB of weights

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Open the model card for the loading code and load the files with transformers.

pip install -U huggingface_hub
hf download ornith-ai/Ornith-1.5-35B-A3B-NVFP4

How to use it: Once the files are on disk, load them with the transformers library and give the model a coding or terminal task. The exact loading code sits on the model card.

Tool | What it does
pip install -U huggingface_hub | Installs the Hugging Face download tool
hf download | Copies the model files to your computer

Example: Inside a terminal agent that runs this model, you ask it to find why a project's tests fail and fix them. The model answers with shell commands and code changes for the agent to apply.

For newcomers and experts: It is a language model tuned for programming and agent work, saved in a smaller file format. For a first try, run the two commands above to fetch the files, then follow the model card to load them. Expect a download of about 19.5 GB.

Ornith-1.5 builds on Ornith-1.0, which started from Qwen3.5 and Gemma4 with extra pretraining and post-training. The new loop keeps producing training tasks, finds ways to solve them and updates the model through reinforcement learning, tuning tasks, scaffolds and solution attempts together. The scores below come from the card for the full model, so a quantized copy may differ.

The model writes new training tasks
It looks for strategies that solve them
Solution attempts earn a reward
Reinforcement learning updates the model

Good to know: The card compares the full model on Terminal-Bench 2.1 with the Terminus-2 setup, where higher is better.

Title: Terminal-Bench 2.1 with Terminus-2, score
Ornith-1.5-35B-A3B: 67.8 points
Ornith-1.0-35B-A3B: 64.2 points
Qwen3.5-397B: 53.5 points
Qwen3.6-35B-A3B: 52.5 points
Muse-Glimmer-30B: 51.7 points
Gemma-4-31B: 42.1 points

• License: MIT
• Cost to run: no API key or paid service is named, you supply the hardware
• Data it reaches: the source names no files, accounts or web access beyond downloading the weights from Hugging Face
• Not for: trusting the card's scores for this copy without testing, since they describe the original model

On Hugging Face, 14 November 2026

• Task: text generation
• Library: transformers
• Parameters: 19.5B
• Weights: about 19.5 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 1,124,337
• Likes: 73
• License: mit
• Quantized versions: 2
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, text-generation, conversational, 8-bit, modelopt

Links

• Model: https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B-NVFP4
• Files: https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B-NVFP4/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ornith-ai/Ornith-1.5-35B-A3B-NVFP4

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