# Outlier

> Mac-native local AI runtime with a custom paged Mixture-of-
> Experts inference engine. Runs frontier-scale models like
> Qwen3.5-397B-A17B on 64GB Apple Silicon Macs by streaming
> expert weights off SSD. Offline, no API tokens, no usage caps.
> Includes a Claude-Code-class agent loop (plan / act / verify /
> recover), MCP connectors, Projects, Code Agent v2, and 7 model
> tiers from 4B to 397B.

## Key documents
- [Homepage](https://outlier.host/)
- [About — who builds Outlier and why](https://outlier.host/about/)
- [Learn: how local AI works](https://outlier.host/learn/)
- [What is local AI?](https://outlier.host/learn/what-is-local-ai/)
- [Comparisons hub (Outlier vs Claude, Ollama, LM Studio, Jan)](https://outlier.host/vs/)
- [For your work: lawyers, healthcare, finance, teams, and more](https://outlier.host/for/)
- [Best local AI apps and LLMs](https://outlier.host/best/)
- [How-to guides: install, RAM, run offline, run models](https://outlier.host/how-to/)
- [Run frontier models on your Mac](https://outlier.host/run/)
- [Why people switch: caps, subscriptions, privacy, energy](https://outlier.host/learn/hit-your-ai-usage-cap/)
- [Benchmarks and data (raw, sourced)](https://outlier.host/data/)
- [Can a local 27B model code like Claude? (measured: ~45% SWE-bench Verified blind, the 3 walls, the execution-oracle lever)](https://outlier.host/data/can-a-local-27b-model-code-like-claude/)
- [Outlier vs Claude - 54-prompt benchmark](https://outlier.host/data/outlier-vs-claude-54-prompt-benchmark/)
- [How an AI agent checks its own work (test-vs-source arbitration, anti-gaming guard)](https://outlier.host/learn/how-an-ai-agent-checks-its-own-work/)
- [GitHub](https://github.com/Outlier-host)
- [Hugging Face](https://huggingface.co/Outlier-Ai)

## Pricing
- Free: Nano 4B + Lite 9B, no account, no token bill
- Pro: $9/mo - all 7 tiers including Plus 397B, Marathon mode, Computer use, Deep Research, long context to 128K
- Founders Lifetime: $249 one-time - everything Pro includes, forever, including all future model releases. No seat cap and no countdown.