Quick recommendation
Not sure where to start? Here’s the short version:| Your device | Recommended model | Backend |
|---|---|---|
| MacBook Air (8 GB) | Gemma 3n E2B or Qwen3 4B | MLX |
| MacBook Pro (16 GB) | Qwen3 8B | MLX |
| MacBook Pro (32 GB+) | Qwen3 32B | MLX |
| Any Mac (fallback) | Qwen3 4B | llama.cpp |
| Browser only | Qwen3 1.7B or Qwen2.5 0.5B |
Downloading a model
Go to the AI section
Find the model management panel where you can see available and downloaded models.
Select a model
Choose a model from the list. The UI shows the model size, memory requirement, and compatible backends.
Understanding model sizes
Larger models are generally smarter but require more memory and respond more slowly. Here’s what to expect:| Size class | Parameters | Good for |
|---|---|---|
| Small (0.5B - 1.7B) | Fast, lightweight | Quick answers, simple tasks, low-memory devices |
| Medium (4B - 8B) | Balanced | General use, research, writing, code |
| Large (12B - 32B) | More capable | Complex reasoning, detailed analysis, nuanced writing |
| Very large (70B+) | Most capable | Advanced reasoning, demanding tasks (requires 64+ GB) |
Choosing a backend
The backend determines how the model runs on your hardware. See On-Device AI for the full technical breakdown.| Backend | Platform | Speed | Model support |
|---|---|---|---|
| MLX | macOS (Apple Silicon) | Fastest | Up to 235B parameters |
| llama.cpp | macOS (desktop app) | Moderate | Up to 70B parameters |
| WebGPU | Any browser | Varies by GPU | Up to 1.7B parameters |
Managing your model library
Switching models
You can switch between downloaded models at any time from the Oasis settings or the model selector. The current model unloads and the new one loads in its place.Storage
Models are stored locally on your device. A small model (0.5B) takes around 500 MB of disk space, while a large model (32B) can take 20+ GB. Check your available disk space before downloading larger models.Removing models
If you need to free up space, you can delete downloaded models from the model management panel. They can always be re-downloaded later.LoRA adapters
After choosing a base model, you can optionally attach a adapter through your persona configuration. Adapters specialize the model for specific tasks without requiring a separate download of a full model.Learn more
On-Device AI
Technical deep dive into backends, tiers, and hardware routing.
Oasis
The app where you interact with your local AI.