Outlier vs Jan — two takes on local AI for Mac

Both of these put a model on your own machine, no cloud account, Wi-Fi optional. But they're built for different people. Jan is the open-source, hack-on-it, runs-anywhere option that lets you drop in any GGUF you like. Outlier is the Mac-native, batteries-included one whose whole reason to exist is running a model bigger than the RAM you have. Here's where each one fits.

What Jan is

Jan is an open-source local AI desktop app, and a genuinely good one. It's free, it's MIT-licensed, and it runs on Windows, Mac, and Linux. Under the hood it loads GGUF models through a llama.cpp engine, so the community model library is right there for you. You get a clean chat GUI, you can swap models in and out, and it exposes an OpenAI-compatible local API so other tools can talk to it. Because it's open source, you can read exactly what it does with your data, which is the answer to a question a lot of cloud apps would rather you didn't ask.

If your priorities are "free," "I want to see the source," and "it has to run on my Linux box and my Mac," Jan is a strong pick. None of that is a backhanded compliment. A lot of people want exactly that, and Jan does it well.

What Outlier is

Outlier is a Mac desktop app for Apple Silicon. One signed download, no account, no terminal, no Docker, and it works with Wi-Fi off. The inference backend is built on Apple's MLX framework, and the models are open-weight, published on HuggingFace. What it adds on top of chat is a full app: a coding agent, deep research, vision, computer use, voice, and project memory, rather than a chat window alone.

The piece that defines it is the inference engine. Outlier's patent-pending paged engine streams individual expert tensors off the SSD as a model needs them, instead of holding the whole thing in memory. That's how a 397B-parameter model runs at roughly 11 GB peak RSS on a 64 GB Mac. The free tier covers Nano and Lite, local and private with no account; the paid tier opens the bigger tiers up to Plus 397B.

The core difference

Two differences, really, and they're the whole story.

RAM-bound vs paged streaming. Jan, like other GGUF tools, loads the model into memory to run it. That's a sane, fast design, and it means the largest model you can run is the one that fits in your RAM. On a 64 GB Mac, once you leave room for the OS, that's roughly 50 GB of weights. Outlier takes the other road: it pages experts off disk, so the model on screen can be far larger than the memory in the machine. The 209 GB of weights behind Plus 397B run at ~11 GB peak RSS because most of those experts sit on the SSD until they're actually called. Different trade — streaming a giant model is slower per token than a model living in RAM — but it's the trade that lets the giant model run at all.

Chat GUI vs full app. Jan is, at its core, a clean front end for chatting with a model you supply. Outlier ships the model lineup and wraps it in an agent that can edit files, do multi-step research, read images, and drive your screen. Neither is automatically better. If you already have your GGUF picked out and you just want to talk to it, Jan's focus is a feature. If you want the whole workflow handed to you on a Mac, that's Outlier.

Comparison table

Dimension	Jan	Outlier
Platforms	Mac, Linux, Windows	Mac (Apple Silicon, M1+)
License	Open source (MIT)	Proprietary app, free Nano+Lite tier
Format / engine	GGUF (llama.cpp)	MLX, paged expert streaming
Models > your RAM	Bound by RAM (won't fit)	Plus 397B at ~11 GB peak RSS
Bring your own model	Yes, any GGUF	Curated tiers + custom MLX models
Built-in agent / research / vision	Chat GUI + local API	Agent, deep research, vision, computer use
Local OpenAI-compatible API	Yes	Yes
Pricing	$0, open source	$0 Free + Pro $20/mo or $149/yr + lifetime from $99

Who should pick which

Pick Jan if you want an open-source app you can audit, you're on Windows or Linux (or want one app across all three), you like bringing your own GGUF models, and the models you run already fit in your machine's memory. It's free, it's transparent, and it stays out of your way.

Pick Outlier if you're on an Apple Silicon Mac, you want a finished app rather than a chat window to wire up, and you want the option to run a model bigger than your RAM — a 397B on a 64 GB Mac — plus the coding agent, deep research, and vision that come with it. It's the local AI you stop renting from the cloud for most daily work.

And these aren't mutually exclusive in spirit. If you've got a Linux server and a Mac, it's perfectly reasonable to keep Jan for the cross-platform, open-source side and use Outlier on the Mac when you want the big model and the built-in agent.

Frequently asked questions