Choose a model
Ekorbia doesn’t ship with a model baked in — you pick one and it downloads. With the bundled engine (the default), that’s a couple of clicks inside the app; no terminal, no separate install.
The easiest path: guided setup
The first time you launch Ekorbia with no model installed, it offers to set one up for you. It detects how much memory your Mac has, recommends a right-sized Gemma 4 model from the built-in catalog, and downloads it (plus the small embedding model, if you want folder search) with a progress bar. You’ll land in a working chat in a couple of minutes.
Want a different model than the one suggested? Click Choose a different model to open the full model manager, or read on.
The built-in catalog
Open Settings → Models (or click the model picker in the bottom-right of the composer and choose Manage models…) to see the catalog. Each entry shows its download size and how much memory it wants, with the recommended pick flagged. Click Download and a progress bar tracks it; the model appears in the picker the moment it’s ready.
Every catalog model is a Gemma 4 build and can see images:
| Model | Download | Wants | Good for |
|---|---|---|---|
| Gemma 4 E2B | ~4.3 GB | 8 GB RAM | Fast responses on any Apple-Silicon Mac |
| Gemma 4 E4B | ~6.1 GB | 8 GB RAM | A balanced all-rounder |
| Gemma 4 12B | ~7.2 GB | 16 GB RAM | Stronger reasoning and writing (recommended) |
| Gemma 4 26B-A4B | ~15.4 GB | 32 GB RAM | Long-form, complex analysis |
Not sure? Take the guided recommendation — it’s sized to leave headroom for the rest of your apps. You can download as many as you like; they live side-by-side, and you switch between them with the model picker in the bottom-right of the composer (see The chat window).
Downloads are checksummed and resumable: closing the manager keeps them running, and Cancel keeps what’s already downloaded so retrying later picks up where it left off. See Managing models for the full tour, including deleting models.
The embedding model (for attachments and search)
To attach files or folders and have Ekorbia find relevant chunks when you ask questions, you also need a small embedding model. The guided setup offers to grab it for you; otherwise download nomic-embed-text (about 274 MB) from the catalog.
You only need this if you plan to attach files or folders. Pure text chats don’t use it.
Choosing a backend
Most people should stay on the bundled engine — it’s the default and needs nothing extra. But Ekorbia can run models three ways, switchable any time under Settings → Backend:
- Bundled engine (default) — Ekorbia runs models itself from the catalog above. Easiest, nothing to install.
- Ollama — if you already use Ollama or prefer it. Install it, pull a model with
ollama pull <name>, then choose Ollama under Settings → Backend. Ekorbia’s model picker then lists your Ollama models. - Custom endpoint — any OpenAI-compatible server (LM Studio, llama.cpp’s
llama-server, vLLM, …). Enter its URL under Settings → Backend; that server owns its own models.
Switching backends is safe — nothing is lost. Because different backends produce different embeddings, if you’ve indexed attachments and then switch, Ekorbia offers a one-click re-index so search stays accurate.