Ollama Setup Guide
Install Ollama and connect it to Maexry as a Bridge engine. Run AI on your own hardware — your data never leaves your machine.
Ollama Setup Guide
Run AI classification entirely on your own hardware. With Ollama as your Bridge engine, your emails are processed locally — nothing is sent to any cloud service.
What Is Ollama?
Ollama is a local LLM runtime that runs open-source models directly on your machine. It handles model downloading, GPU acceleration, and serves a simple API that Maexry connects to.
Why it matters for Maexry: sovereign AI. With Ollama (Bridge), classification runs entirely on your own machine — your email never leaves your device for AI processing. No API keys to manage, no usage fees, no third-party data processing. Just your hardware, your models, your rules.
Installation
macOS
Install via Homebrew or download the app:
brew install ollama
Or download the macOS app from ollama.com/download.
Linux
Run the install script:
curl -fsSL https://ollama.com/install.sh | sh
Windows
Download the installer from ollama.com/download and follow the setup wizard.
Start the Server
After installing, start Ollama:
ollama serve
Ollama runs on http://localhost:11434 by default. Keep it running in the background while using Maexry.
Network Access
By default, Ollama only listens on localhost. If you want to connect from another device on your network (e.g., Maexry on your phone connecting to Ollama on your desktop), you need to expose it:
OLLAMA_HOST=0.0.0.0 ollama serve
Find your machine's local IP address:
- macOS / Linux:
ifconfigorip addr - Windows:
ipconfig
Your Ollama URL from other devices will be http://<your-ip>:11434.
Recommended Models
Based on our benchmarks on an NVIDIA RTX 4090 (24GB VRAM):
| Model | Size | Speed | Best For |
|---|---|---|---|
gemma4:e4b | 9.6 GB | ~162 tok/s | Fast classification, daily email triage |
gemma4:26b-a4b-it-q4_K_M | ~16 GB | Similar to 4B (MoE) | Best quality/speed ratio — 26B knowledge at 4B speed |
gemma4:26b | ~18 GB | ~40–60 tok/s | Highest quality, needs 24GB+ VRAM |
To download a model:
ollama pull gemma4:e4b
For most users, gemma4:e4b offers the best balance of speed and accuracy for email classification.
Hardware Recommendations
| Tier | Hardware | What You Can Run |
|---|---|---|
| Minimum | 8 GB RAM, any modern CPU | Small models on CPU (slower, but works) |
| Recommended | NVIDIA GPU with 8 GB+ VRAM (RTX 3060+) or Apple Silicon Mac (M1+) | 8B models at usable speed |
| Optimal | NVIDIA RTX 4090 (24 GB) or Mac Studio M2 Ultra | 26B+ models with fast inference |
Apple Silicon Macs use unified memory — a 16 GB M1 can run 8B models comfortably. A 32 GB M-series Mac handles most models without issue.
Connecting to Maexry
- Open Maexry → Settings → AI Engine.
- Select Bridge (Ollama) as your engine.
- Enter your Ollama URL:
- Same machine:
http://localhost:11434 - Another device on your network:
http://<your-ip>:11434
- Same machine:
- Select your model from the dropdown.
- Tap Test Connection to verify.
Once connected, Maexry will use your local Ollama instance for all AI classification and processing.
Performance Tips
- First inference is slow. Ollama loads the model into memory on the first request. Subsequent calls are fast.
- Keep Ollama running in the background for instant responses. The model stays loaded in memory.
- For classification (short prompts and responses), even 8B models perform well. You don't need the largest model.
- Larger models shine for category generation and email summarization, where nuance matters more.
- Monitor VRAM usage. If your GPU runs out of memory, Ollama falls back to CPU, which is significantly slower. Choose a model that fits your hardware.