Noema

Noema is an AI assistant built for Apple devices that runs large language models, file search, and tool integrations entirely on your hardware. It works on iPhone, iPad, Mac, and Vision Pro without requiring a cloud connection for basic chat. Your conversations stay on device unless you explicitly connect to a remote endpoint, and the app makes those network paths visible so you know exactly what's local and what isn't.

The pitch is privacy through architecture rather than promises. Most AI assistants send every prompt to a server, which means your data passes through someone else's infrastructure. Noema flips that by running inference locally using a unified library that supports GGUF, MLX, Core ML, ExecuTorch, and Apple Foundation Models. You can load open models, run them offline, and keep everything on hardware you control.

Beyond chat, it includes on device PDF reading and dataset grounding. You can point it at your own files and have it retrieve relevant context without uploading anything. There's also web evidence access for when you want to pull in citations from the internet, MCP tool connections for integrating external utilities, and an off grid mode that blocks all HTTP and HTTPS connections if you want absolute network isolation.

Noema ships with its own 2 billion parameter model, Noema 2B, tuned for reasoning, math, and instruction following. It's designed to run well on Apple hardware without eating your battery or requiring top tier specs. You can also connect to external endpoints like LM Studio, Ollama, OpenAI compatible APIs, or Noema Relay for cases where you want to access larger models remotely but still manage everything through one interface.

The app is free with no subscription. There's no cloud cost because there's no cloud requirement for local chat. You download it from the App Store, load a model, and start using it. Remote endpoint features are optional and visibly separated from the local experience.

It fits anyone in the Apple ecosystem who wants an AI assistant without handing their prompts to a third party. Privacy conscious professionals, developers experimenting with local models, researchers who need file grounding without data leaving their machine, or anyone who's uncomfortable with the default assumption that every AI query goes to a server. Noema treats local first as the default, not an afterthought.

Vision Pro support extends the same capabilities to spatial computing. Whether you're on a phone, tablet, laptop, or headset, the experience stays consistent and the privacy model stays the same. If you've been looking for a way to run AI tools on Apple hardware without trusting a cloud provider, this is one of the more complete implementations available.