What Is Ornith 1.5 and How Does It Work?

Ornith 1.5 is a new family of open-source AI models that can write code, fix bugs, and run tasks on their own. It comes in three sizes: 9B, 35B, and 397B parameters. The biggest model scores close to Claude Opus 4.8 on real coding tests, but the real story is how it trains itself.

Instead of humans writing every training task, Ornith 1.5 creates its own tasks, builds tools to solve them, solves them, and then uses the results to get better. It repeats this cycle, making each task harder. This is a completely different way to train AI, and it means the model is always improving itself without human help.

This matters because it makes the AI more useful for real business tasks like debugging code or automating workflows.

Key Features of Ornith 1.5

Here are the main features that make Ornith 1.5 stand out:

  • Three sizes: 9B (small), 35B (medium), and 397B (large) parameters, so you can choose based on your needs.
  • Self-improving loop: The model writes its own tasks, plans, and tools, then scores itself and creates harder tasks.
  • Scoring system: It checks validity, difficulty, and novelty to ensure tasks are real, challenging, and not repetitive.
  • MIT license: You can use it commercially and build on it.
  • Compatibility: Works with tools like Olima and llama.cpp.
  • Phone capability: A squeezed version of the 9B model can run on iPhone and Android.

These features make it accessible for small developers and businesses, not just big tech companies.

How to Get Started with Ornith 1.5

Getting started with Ornith 1.5 is straightforward because it is open source and works with common tools. Here are the steps:

  1. Choose a size: 9B for phones, 35B for daily use, or 397B for maximum power.
  2. Download the model from the official release page or a platform like Olima.
  3. Install compatible tools like llama.cpp on your machine.
  4. Load the model into your chosen environment.
  5. Give it a task, such as fixing a bug or building a form.
  6. Let it plan, execute, and check its own work.
  7. Review the results and adjust as needed.

Since it is MIT-licensed, you can integrate it into your own apps and workflows without restrictions.

Start with the 35B model for a balance of speed and performance.

How to Use Ornith 1.5 in Real Life

You can use Ornith 1.5 for practical business tasks. In the episode, Julian gave examples like:

  • Feeding it your website's code and asking it to find broken links, fix them, and confirm they're fixed.
  • Building a lead capture form, adding basic security, writing tests, running them, and fixing anything that fails.
  • Automating lead capture analysis: "look at our lead capture system, find where leads are dropping off, write a fix, test it, and tell me what changed."

These are multi-step tasks that the model can handle from start to finish, not just autocomplete.

You can also use it for coding, debugging, and automation. The benchmarks show it is an actual agent that does a job, not just talk.

For a small team, the 35B model is fast and light, making it practical for daily use.

Benchmark Results and Comparisons

The episode showed several benchmark results. Here are the key numbers:

  • Terminal Bench 2.1: Ornith 397B scored 86.1, Claude Opus 4.8 scored 85.0.
  • SWE Dash Bench Verified: Ornith 397B scored 86.0, Claude scored 85.8.
  • 35B model: scored 68.5 on Terminal Bench, 79 on SWE Dash, and 89.2 on GPQA Diamond.
  • 9B model: scored 70.6 on SWE Dash and 86.4 on GPQA Diamond.

The big model is close to Claude but not better at everything. The medium and small models are impressive for their size.

These are Ornith's own published results, so they may not hold up in every real-world task. The community is testing them.

Tips, Limits, and Mistakes to Avoid

Be careful when using Ornith 1.5. The benchmarks are from Ornith, so they might not match your experience. Different setups and tools can change results.

  • Test yourself: Don't build your entire workflow around it without testing first.
  • Be realistic: It may not beat Claude in every task.
  • Understand the 20% rule: The training aims for tasks the model solves 20% of the time, so expect some failures.
  • Check novelty: The system avoids repetitive tasks, but you should too for real work.

Also, the model is not a magic solution. It needs clear instructions to work well.

Avoid expecting it to handle everything perfectly from day one. Start with simple tasks and scale up.

Who Should Use Ornith 1.5 and Why It Matters

Ornith 1.5 is for developers, small businesses, and agencies that want powerful AI without huge costs. Since it's open source, you can run it yourself and customize it.

It matters because open models are closing the gap with the big frontier models. An open model near Claude Opus 4.8 is a big deal.

Small models are also getting capable enough to run on phones, which opens up new possibilities.

The biggest reason is that training is becoming automated. If this keeps working, future open models will be better at improving themselves.

This could change how you use AI in your business, making automation more accessible.