Local Models with Ollama

Run Omnideck fully offline using local models via Ollama.

If you'd rather run models locally than send data to a cloud provider, Omnideck supports Ollama for fully local inference. Install Ollama on your host, pull the models you want, and Omnideck lists them automatically in the setup wizard.

Install Ollama

Download and install Ollama from ollama.com. After installing, pull the models you want to use:

ollama pull kimi-k2.5      # main model (chat, compaction, titles)
ollama pull qwen3.5        # vision model (optional — for image input)

Verify Ollama is running and your models are available:

ollama list

Recommended models

Role Suggested model Notes
Main kimi-k2.5 Used for chat, context compaction, and conversation title generation
Vision qwen3.5 Must support image input — or use your main model if it does

The main model handles everything by default. A separate vision model is only needed if you want a model specifically optimized for image input.

Connect Omnideck to Ollama

During omnideck install, the wizard checks whether Ollama is reachable on the host and warns if it isn't. If Ollama is running, select it as your provider in the setup wizard — Omnideck will list the models you've pulled.

If Omnideck is already running, go to Settings → Providers in the web UI and add Ollama as a provider. The list of available models updates automatically from your local Ollama install.

macOS and Docker Desktop

On macOS with Docker Desktop, Ollama must be bound to 0.0.0.0 (not just 127.0.0.1, which is the default) so the container can reach it on the host. Set this before starting Ollama:

export OLLAMA_HOST=0.0.0.0
ollama serve

Or set it permanently in your shell profile. omnideck doctor will report whether Ollama is reachable from inside the container.

Ollama cloud variants

Ollama can also broker cloud-hosted model variants alongside your local models, so they appear the same way in Omnideck. After ollama login, pull a cloud variant:

ollama pull kimi-k2.5:cloud

Cloud variants skip your local GPU but still need to be pulled so Ollama exposes them. This is useful if you want a unified provider interface that can mix local and cloud models.

Troubleshooting

"Ollama connection refused" in omnideck doctor

Only matters if you're using local models. Make sure:

  1. Ollama is running (ollama serve)
  2. On macOS / Docker Desktop, OLLAMA_HOST=0.0.0.0 is set (see above)
  3. On Linux with Podman, the firewall isn't blocking the host network bridge

Models don't appear in the setup wizard

The wizard lists models already pulled in Ollama. Pull your desired models first (ollama pull <model>), then reopen Settings in the Omnideck UI.