Outlier vs PrivateGPT: a Mac app vs a self-hosted API layer

I'm Matt Kerr — I build Outlier, so I have an obvious stake here. I'm writing this anyway because the two get lumped together in search results while solving genuinely different problems. Both are private. Both are local. That's where the similarity ends: PrivateGPT is a developer platform with 57,000 GitHub stars behind it; Outlier is a consumer Mac app.

What PrivateGPT actually is in 2026

The canonical project is the GitHub repo zylon-ai/private-gpt, started in May 2023 and licensed under Apache 2.0. As of 23 July 2026 it has 57,354 stars and 7,607 forks. It describes itself as a complete API layer for private AI applications running on local models: RAG, skills, tools, MCP, text-to-SQL, and more.

The important detail — the thing most old tutorials get wrong — is that PrivateGPT does not run models. Version 1.0.0, shipped 3 June 2026, was a ground-up rewrite that moved model execution out of the project. You set OPENAI_API_BASE to point at an OpenAI-compatible server you run yourself; Ollama, LM Studio, llama.cpp and vLLM are named explicitly. v1.0.1 followed on 18 June 2026 with bug fixes and prompt-caching work. Before that came v0.6.2 in August 2024, and pre-1.0 PrivateGPT did run models itself — so nearly every walkthrough written between 2023 and mid-2026 describes software that no longer works that way. Check the date on any tutorial.

Installation is command-line everywhere: Homebrew on macOS, uv tool install on Linux and Windows, or Docker. The uv path pins Python 3.11 exactly — the docs say 3.10 and 3.12+ aren't supported. Run private-gpt serve and you get a browser-based "Workbench" UI on port 8080 alongside the API. The docs are candid that it's a local demonstrator: no access control, chats kept in browser localStorage, and not every backend capability exposed. It's a developer console, not a product.

One more trap: an unrelated commercial PrivateGPT is sold on AWS Marketplace by BigCheese with Abstracta, deployed into your own AWS account at custom pricing. No documented relationship to the Zylon project.

What Outlier is

Outlier is a native macOS app for Apple Silicon (M1 or later, macOS 12+). You download it, pick a tier, and the model runs on your machine. No account, no usage caps, no terminal. It ships seven tiers: Nano 4B (2.37 GB, 6 GB RAM, 71.7 tok/s on M1 Ultra), Lite 9B (5.04 GB, 12 GB), Quick 26B-a4b (15.61 GB, 16 GB), Core 27B and Code 27B (15.13 GB, 24 GB, 20.7 tok/s), Vision 35B-a3b (19.0 GB, 24 GB), and Plus 397B-a17b (209 GB, 64 GB, 1.59 tok/s, streaming experts from SSD at ~11 GB peak RSS).

Nano and Lite are free with no account. Pro covers all seven tiers, with Founders Lifetime at $249. The base models are Qwen and Gemma derivatives — six tiers Apache 2.0, Quick under the Gemma Terms of Use — and the weights are published at huggingface.co/Outlier-Ai. The app itself is not open source. That's a real difference from PrivateGPT and I'll come back to it.

What each one hands you out of the box

Where PrivateGPT genuinely wins

Several of these aren't close, so I'll say them flatly.

• It's open source, and that's a stronger privacy guarantee than mine. Apache 2.0 means every claim about where your data goes is verifiable in the source. Outlier's privacy claims you take on trust from me. However local a closed app is, it can't match auditability.
• Cross-platform. macOS, Linux and Windows all get first-class install paths, plus Docker. Outlier is Apple Silicon only — no Intel, no Windows, no Linux, no mobile. If your team isn't all on modern Macs, Outlier is out of the running and PrivateGPT isn't.
• Backend-agnostic. It works with any OpenAI-compatible server and therefore any model those servers can host. No lock-in to one runtime, quantization format or model family. An app that ships its own engine can't offer that.
• Production primitives. Document ingestion, retrieval with citations, agentic RAG, embeddings, custom tools, MCP connectors, text-to-SQL, tabular analysis, code execution in your workspace. Client access is covered too: the 1.0 docs' "Client libraries" page tells you to install the Anthropic SDK (pip install anthropic or npm install @anthropic-ai/sdk) or LangChain and point it at your own PrivateGPT URL, because PrivateGPT's own endpoint speaks the Claude API even though it talks OpenAI-compatible on the way out to your inference server — so any Claude-compatible client works unmodified. The docs name Claude Code and Cursor as AI clients you can point at PrivateGPT's MCP server. Ignore the first-party pgpt-python and privategpt-sdk-node packages you'll find in search results: both are stuck at 0.1.2 from 2024, target the pre-1.0 API that v1.0.0's release notes call not forward-compatible, and are no longer referenced by the docs.
• It deploys for a team. One server, many clients, on-prem — a single-user desktop app can't do that at all. Read the fine print, though: upstream authenticates everyone with one shared server secret and has no per-user isolation, the docs warn that the Workbench has no access control and should stay local, and they list auth providers (OAuth, LDAP, SSO), multi-tenancy and per-user vector stores as things community forks add rather than things the project ships. Real multi-user deployment is fork work or Zylon's commercial product.
• Free at any scale — no seats, no tokens, no metering — with a funded company (Zylon) maintaining it.

Where Outlier fits better

Outlier's advantage is narrow and specific: it's a finished product for one platform. One thing to install instead of two, no Python version to pin, no embedding model to choose, and a published RAM figure for every tier so you know before downloading whether it'll run on your 16 GB Mac. PrivateGPT publishes only example-model sizes — roughly 24 GB for its reference LLM under Ollama, about 18 GB as a Q4 GGUF, 670 MB for embeddings — not per-machine RAM floors. That's reasonable, since performance comes from whatever server you point it at, but "will this run well on my laptop?" still has no documented answer.

On quality, the only numbers I can show are for my own models. Core 27B matched Claude Opus on 98.9% of rubric checks across a 54-prompt head-to-head, and 100% on the nine hardest prompts; on a blind slice of SWE-bench Verified the local 27B measured about 45% (18/40). None of that says anything about PrivateGPT, which inherits whatever quality your chosen model has — that's the point of it.

Two more honest notes. "Private" in PrivateGPT describes a configuration, not an enforced property: OPENAI_API_BASE accepts a remote endpoint as happily as localhost, so a misconfigured deployment can quietly ship prompts to a third party. And its built-in web tools break locality by design — web search routes queries to Brave's API and needs a Brave key. Outlier's web research is Pro-only and opt-in, and its free tier has none; it also has no image generation, voice mode or cloud sync.

Developer backend or desktop app?

Pick PrivateGPT if you're building on top of private inference rather than just chatting with it, you need RAG over your own documents with citations, you're on Linux or Windows, you want to audit the code, or you need one server behind an internal app — accepting that per-user accounts and isolation are something you or a fork add on top.

Pick Outlier if you're on an Apple Silicon Mac, you want to be talking to a capable local model within a couple of minutes, and you'd rather not maintain an inference server as a second moving part. It's a product, not a framework.

Running both is reasonable. They don't compete for the same slot — one is your desktop chat window, the other is the backend your app talks to.