How to Run Ornith 1.5 on Ollama: Local Setup Guide (2026)

What's New in Ornith 1.5 and Which Tag to Run

Ornith 1.5 comes from DeepReinforce, the same lab behind Ornith 1.0, still published on Hugging Face under the `deepreinforce-ai` organization. The core idea behind the family hasn't changed: train a model on multi-step agentic coding work by rewarding it on whether its self-generated approach actually solves the problem, not on imitating a fixed human-written harness.

What's different in 1.5 is the scope of that self-improvement loop. Ornith 1.0 generated its own scaffold for a given task, then solved it inside that scaffold. Ornith 1.5 generates the training tasks themselves too, alongside the scaffold and the solution, all optimized jointly through reinforcement learning. DeepReinforce describes the model as continuously proposing new problems, discovering strategies for solving them, and improving its own policy from the results, rather than staying bound to a fixed, human-curated task set.

Two capability changes carry over into how you actually use the model. Every Ornith 1.5 tag now accepts image input alongside text. And Ollama's library now hosts the full 397 billion parameter mixture-of-experts checkpoint, sitting alongside the same 9B and 35B sizes from before. The 35B tag activates roughly 3B parameters per token, which is part of why it stays practical on a single high-end consumer GPU despite its larger total size.

Here's what Ollama hosts as of Ornith 1.5's release:

All three tags keep the same 256K token context window as Ornith 1.0. The 397B size is realistically a rented multi-GPU cloud instance rather than a personal machine, covered in the hardware section below.

Install Ollama and Run Ornith 1.5 Locally

Running Ornith 1.5 through Ollama takes the same four steps as any Ollama model: install Ollama, pull a tag, run it, and verify it loaded.

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 1.5

ollama run ornith-1.5

This pulls the default tag, `ornith-1.5:9b`, and drops you into a prompt once the 6.6 GB download finishes:

pulling manifest
pulling 4b7f2e91... 100% ▕████████████████▏ 6.6 GB
success
>>> Send a message (/? for help)

Step 3: Pull a Specific Size

ollama pull ornith-1.5:9b
ollama pull ornith-1.5:35b
ollama pull ornith-1.5:397b

Step 4: Verify the Models

ollama list

NAME                SIZE      MODIFIED
ornith-1.5:9b       6.6 GB    2 minutes ago
ornith-1.5:35b      23 GB     5 minutes ago

Step 5: Test an Agentic Coding Prompt

>>> Write a Python function that validates a JSON schema against a nested dictionary, then propose your own test cases for it.

Because Ornith 1.5 was trained to generate its own tasks and scaffolds, not just solve fixed ones, prompts that ask it to propose its own verification approach tend to surface that training more clearly than a plain code-completion request.

Choosing Between the 9B, 35B, and 397B Tags

Three sizes now means three real hardware tiers instead of two.

`ornith-1.5:9b` is the right starting point on ordinary hardware, the same as it was for 1.0. It scores noticeably higher than Ornith 1.0 on most of DeepReinforce's published benchmarks (see the benchmark section below) while keeping the same modest download size.

`ornith-1.5:35b` is the practical ceiling for a single consumer GPU. A 24 GB card like an RTX 4090 handles it with full offload. Without one, plan on 32 GB or more of system RAM and slower CPU-assisted inference.

`ornith-1.5:397b` is a different category of hardware requirement entirely. At 242 GB, it doesn't fit on any single consumer GPU, and running it well means a multi-GPU cloud instance, not a personal machine.

If you're unsure where to start, pull `ornith-1.5:9b` first. It's a fraction of the download time of the larger tags and gives you a direct feel for what changed since Ornith 1.0 before deciding whether 35B or 397B is worth the extra hardware.

Using Ornith 1.5's New Vision Input

Ornith 1.0 was text-only. Every tag in Ornith 1.5 now accepts images alongside text, which matters for agentic coding work that involves reading a screenshot, a diagram, or a rendered UI rather than only raw code.

Send an Image from the Terminal

ollama run ornith-1.5 "What's wrong with this layout? ./screenshot.png"

The model reads the image directly alongside your text prompt in the same request, useful for debugging a UI bug from a screenshot or checking a rendered diagram against what the code is supposed to produce.

Send an Image via the API

curl http://localhost:11434/api/chat -d '{
  "model": "ornith-1.5:9b",
  "messages": [
    { "role": "user", "content": "Does this chart match the data in report.csv?", "images": ["<base64-encoded-image>"] }
  ],
  "stream": false
}'

How Ornith 1.5 Compares: Real Benchmark Numbers

DeepReinforce published a 12-benchmark comparison for the 397B size of Ornith 1.5 against Ornith 1.0, DeepSeek-V4-Flash-0731 (284B), GLM-5.2 (753B), and Claude Opus 4.8. These are vendor-reported numbers, so treat them as a starting point rather than a final verdict, but the gap over Ornith 1.0 specifically is large enough across every benchmark shown to be worth noting.

A few numbers stand out. On DeepSWE, Ornith 1.5 scores 56.0 against Ornith 1.0's 8.0, a jump too large to attribute to noise alone. Toolathlon-Verified moves from 43.2 to 71.2 across the same generation gap. On WideSearch, Ornith 1.5 actually outscores Claude Opus 4.8, 80.8 to 72.9, despite Opus being a proprietary frontier model.

GLM-5.2 carries almost double Ornith 1.5's parameter count at 753B, yet Ornith 1.5 beats it on 10 of the 12 benchmarks shown here. Parameter count alone doesn't predict which model wins a given benchmark.

Using Ornith 1.5 with OpenClaw and Other Coding Agents

DeepReinforce's own benchmark suite includes MCP-Atlas and Toolathlon-Verified, both built around tool-calling agent behavior, which makes Ornith 1.5 as natural a fit for OpenClaw with Ollama as its predecessor was. Since Ollama exposes the same OpenAI-compatible endpoint for every model it serves, pointing an existing agent at Ornith 1.5 is a configuration change, not a new install.

Point OpenClaw at Ornith 1.5

model:
  default: ornith-1.5:35b
  provider: custom
  base_url: http://localhost:11434/v1
  context_length: 256000

Swap in `ornith-1.5:9b` for lighter hardware, or `ornith-1.5:397b` if you're running it through a rented multi-GPU instance.

Call Ornith 1.5 Directly from the API

curl http://localhost:11434/api/chat -d '{
  "model": "ornith-1.5:9b",
  "messages": [
    { "role": "user", "content": "Add input validation to this function and explain the test cases you would write for it." }
  ],
  "stream": false
}'

{
  "model": "ornith-1.5:9b",
  "message": {
    "role": "assistant",
    "content": "Here is the validated function, along with the edge cases I would test..."
  },
  "done": true
}

Troubleshooting

Frequently Asked Questions