How to Run Ornith 1.0 on Ollama: Local Setup Guide (2026)

What Ornith 1.0 Is and How Its Self-Scaffolding Training Works

Ornith 1.0 comes from DeepReinforce, a research lab focused on reinforcement learning for coding agents, with weights published on Hugging Face under the `deepreinforce-ai` organization. The model targets agentic coding specifically: multi-step tasks where an AI has to plan, call tools, run tests, and revise its own work, rather than answer a single prompt.

The training loop behind the self-scaffolding idea works in rounds. DeepReinforce samples candidate scaffolds for a given coding task, runs the resulting trajectories against tests or benchmark harnesses, and scores each attempt on whether the code actually passes, not just whether it looks plausible. The model then gets reinforced on the scaffold-and-solution pairs that worked, so over many rounds it learns which kinds of self-generated harnesses tend to lead somewhere useful. According to DeepReinforce's model card, this is what lets the model "discover better search trajectories and generate higher-quality solutions" compared to filling in a fixed, human-written harness.

Ornith's smaller sizes are post-trained on top of Gemma 4, and its larger sizes build on Qwen 3.5, which explains why the family spans such a wide parameter range while sharing the same training recipe. Every size ships under an MIT license, the same permissive terms as the base models it builds on.

DeepReinforce reports Ornith's results on Terminal-Bench 2.1, SWE-Bench, NL2Repo, and OpenClaw, benchmarks chosen specifically because they test multi-step, tool-using coding work rather than single-turn code completion. The company's own numbers put Ornith ahead of comparably sized open-weight coding models on these benchmarks, though independent third-party benchmark runs were not yet published in the days after launch. Treat vendor-reported scores as a starting point, not a final verdict.

Here's what Ollama hosts as of late June 2026:

DeepReinforce's Hugging Face collection also includes 31B dense and 397B mixture-of-experts checkpoints, but neither tag has reached Ollama's official library yet.

Install Ollama and Run Ornith Locally

Running Ornith through Ollama takes four steps: install Ollama, pull a tag, run the model, and verify it loaded. The whole setup takes under ten minutes once the download finishes.

Step 1: Install Ollama

# Linux and macOS, one-command installer
curl -fsSL https://ollama.com/install.sh | sh

On Windows, download the installer from ollama.com/download, or use winget:

winget install Ollama.Ollama

Verify the installation:

ollama --version
# Expected: ollama version 0.6.x or higher

Step 2: Pull and Run Ornith

ollama run ornith

This pulls the default tag, currently `ornith:9b`, and drops you into a prompt once the 5.6 GB download finishes:

pulling manifest
pulling 3a8e1c9f... 100% ▕████████████████▏ 5.6 GB
success
>>> Send a message (/? for help)

Step 3: Pull a Specific Size

If you want the 9B and 35B tags side by side, pull each explicitly instead of relying on the default:

ollama pull ornith:9b
ollama pull ornith:35b

Step 4: Verify the Models

ollama list

Both tags should appear with their download sizes:

NAME            ID              SIZE      MODIFIED
ornith:9b       7f2a91b3...     5.6 GB    2 minutes ago
ornith:35b      c4d8e027...     21 GB     5 minutes ago

Step 5: Test an Agentic Coding Prompt

>>> Write a Python function that retries a failing HTTP request up to 3 times with exponential backoff, then explain your scaffold for testing it.

Because Ornith was trained to generate its own task scaffold alongside the solution, responses to multi-step prompts like this often include the model's own plan for verifying its answer, not just the code.

Choosing Between ornith:9b and ornith:35b

The right tag depends on your hardware, not just on wanting the bigger model. Here's how the two compare:

`ornith:9b` runs comfortably on a laptop with 16 GB of RAM and no dedicated GPU, though responses are faster with even a modest GPU. It's the right starting point for testing Ornith's self-scaffolding behavior before committing disk space and download time to the larger model.

`ornith:35b` is the size DeepReinforce's own benchmark numbers lean on most heavily for agentic coding work. Running it well, meaning full GPU offload rather than slow CPU fallback, needs a GPU with at least 24 GB of VRAM, like an RTX 4090. Without that, you're looking at 32 GB or more of system RAM and noticeably slower responses as Ollama offloads layers to the CPU.

If you're unsure which to start with, pull `ornith:9b` first. It downloads in a fraction of the time and gives you a feel for how the model's self-generated scaffolds behave before you decide whether the 35B size is worth the extra hardware.

Using Ornith with OpenClaw and Other Coding Agents

DeepReinforce specifically benchmarked Ornith on OpenClaw, which makes it a natural fit for anyone already running OpenClaw with Ollama as a local coding agent. Because Ollama exposes the same OpenAI-compatible endpoint for every model it serves, switching OpenClaw (or any other Ollama-compatible agent) over to Ornith is a configuration change, not a new install.

Point OpenClaw at Ornith

Update the agent's model configuration to use Ornith's tag instead of whatever local model it was pointed at before:

model:
  default: ornith:35b
  provider: custom
  base_url: http://localhost:11434/v1
  context_length: 256000

Use `ornith:9b` in place of `ornith:35b` if you're running on lighter hardware. The rest of the configuration stays the same.

Call Ornith Directly from the API

For scripts or custom tooling rather than a pre-built agent, Ornith responds to standard Ollama API calls:

curl http://localhost:11434/api/chat -d '{
  "model": "ornith:9b",
  "messages": [
    { "role": "user", "content": "Refactor this function to handle null inputs and add a test case." }
  ],
  "stream": false
}'

{
  "model": "ornith:9b",
  "message": {
    "role": "assistant",
    "content": "Here is the refactored function, the null check I added, and a pytest case covering the null-input path..."
  },
  "done": true
}

Troubleshooting

Frequently Asked Questions