Outlier vs text-generation-webui (oobabooga): two local AI apps compared

I'm Matt Kerr. I build Outlier, so I have an obvious stake here — read this with that in mind. Every claim about TextGen is verifiable in its repo, release notes or requirements files. And on several axes, TextGen is the more capable software.

Privacy is not the axis — both of these are local

Most "vs" articles about local AI lean on privacy. That doesn't apply here. TextGen's README describes it as "100% offline and private, with zero telemetry, external resources, or remote update requests," and there's no vendor account behind it, so there's nothing to opt out of. Outlier is the same: on-device, no account, works with Wi-Fi off. Both have opt-in network features that break that posture — TextGen's web search queries DuckDuckGo and --public-api routes traffic through a Cloudflare tunnel. Neither is on by default, but the second is a foot-gun for anyone assuming the app is offline in every configuration.

What TextGen actually is now

First, the name. The project has been renamed: the canonical repo is github.com/oobabooga/textgen and the old URL redirects there. Most people still say "oobabooga," but the release assets say textgen.

It's AGPL-3.0 and entirely free — no account, no subscription, no metering. It started in December 2022 and it's one of the largest projects in this space: 47,482 stars, ~5,980 forks, 836 open issues at the time of writing. v4.7 through v4.9 all shipped inside about three weeks in May 2026; v4.9 added MTP speculative decoding, web-search snippet extraction, and a live tokens-per-second readout. Five inference backends — llama.cpp, ik_llama.cpp, Transformers, ExLlamaV3 and TensorRT-LLM — swap without restarting, and the PC side gets separate portable builds for CUDA 12.4, CUDA 13.1, ROCm 7.2, Vulkan and CPU-only.

What Outlier is

Outlier is a native macOS app for Apple Silicon — M1 or later, macOS 12 or newer. No Intel Macs, no Windows, no Linux, no mobile app. Models run fully on-device, no account, no usage caps.

Seven tiers ship with it, from Nano 4B (2.37 GB, 6 GB RAM, 71.7 tok/s on an M1 Ultra) through Core 27B and Code 27B up to Plus 397B-a17b, which needs 64 GB and streams experts from SSD at about 11 GB peak RSS. Nano and Lite are free; Pro is a one-time purchase for all seven. Weights are published at huggingface.co/Outlier-Ai; the app itself is not open source.

Interface, setup, and what's in the box

This is the real dividing line. TextGen is a Gradio web UI wrapped, since v4.7.3, in a bundled Electron shell, so it opens in its own window rather than a browser tab. It's also a server: run it with --api and you get an OpenAI- and Anthropic-compatible endpoint on port 5000, with streaming and tool calling. There's no interactive chat CLI — the command line is for configuration (--cpu, --split-mode, --spec-type), not conversation. The portable builds don't need installing, but on macOS the release notes tell you to run xattr -cr on the folder first, because the builds aren't notarized.

The bigger difference: TextGen ships with no model. You pick the weights, the quantization and the context settings, usually by pulling a GGUF into user_data/models. That's the source of both its flexibility and its learning curve. Outlier inverts it: choose a tier, it downloads, it runs, and it won't offer you something your machine can't hold.

Platforms and licensing — where Outlier is the narrower product

I'll be blunt, because this is the clearest disadvantage on my side. TextGen runs on Windows, Linux and macOS, Apple Silicon and Intel, with the same workflow everywhere. Outlier runs on Apple Silicon Macs and nowhere else. PC with a good GPU, Linux workstation, Intel MacBook? Outlier isn't an option and TextGen is.

Licensing cuts the same way. TextGen is AGPL-3.0 — read it, fork it, audit it, keep running your exact version forever regardless of what the maintainer does next. Outlier's app is closed source, and that's a legitimate reason to prefer TextGen.

One caveat on TextGen's Mac story: the Apple Silicon requirements file ships one compiled inference wheel, llama.cpp — ExLlamaV3 and TensorRT-LLM are absent, so Mac users get a subset of the headline backend flexibility. The Transformers loader is still there, though: the same file pins transformers alongside peft, accelerate and torchao, which is the loader LoRA needs. And as of mid-2026 the documentation doesn't say whether Metal acceleration is compiled into the prebuilt macOS wheels; unlike the Windows and Linux assets, the Mac ones carry no acceleration label. I won't assert either way without testing.

Side by side

Where text-generation-webui genuinely wins

• It's free and it's yours. AGPL source plus free binaries, no account, no metering, no vendor who can change the terms.
• It runs on everything. Windows, Linux and macOS, same workflow. Outlier covers one platform.
• Any model, day one. The whole Hugging Face ecosystem — niche fine-tunes, uncensored models, brand-new releases the week they drop. Outlier gives you our seven tiers and nothing else.
• It's a server, and a better one. One --api flag gives you an OpenAI- and Anthropic-compatible endpoint, fronting whatever model you loaded, reachable from other machines with --listen. Outlier does ship an OpenAI-compatible endpoint — it's off until you turn it on in Settings › Developer, then it's http://127.0.0.1:8766/v1 with a required API key — but it's localhost-only, no Anthropic route, and it only serves our seven tiers.
• LoRA fine-tuning built in, including CPU-mode training. One sharp edge: per the compatibility doc, LoRA loading and training both need the Transformers loader, not llama.cpp, which is what most local users run. On a Mac that's still a live option — requirements_apple_silicon.txt installs Transformers, peft and accelerate — but it means leaving the GGUF path you probably started on.
• Wider modality coverage. Vision, PDF and .docx attachments, image generation via diffusers, TTS and voice extensions, MCP support. Outlier has vision but no image generation, no voice, no web search.
• It serves the household. One capable machine plus --listen puts every phone and laptop on the LAN into the same model.
• Community depth. 47k stars, an extension ecosystem, a wiki and an active subreddit — somebody has already documented your problem.

One counterweight, not a gotcha: v4.9 restricted CORS to localhost "to prevent drive-by API access" and closed a path-traversal vector, so earlier versions were exposed to a web page quietly calling your local API. Active hardening is a good sign — but you run the server.

Where Outlier fits better

Outlier's argument is narrow: one signed, notarized Mac app where the model, settings and interface arrive together, with published numbers for what you get. Core 27B matched Claude Opus on 98.9% of rubric checks across a 54-prompt head-to-head, and 100% on the nine hardest; the Code tier scores 0.866 on HumanEval. TextGen can't publish numbers like that — it's model-agnostic by design, so its quality is whatever you loaded, which is the whole point of it.

The other thing you get is the absence of decisions: no quantization, no flags, no Terminal. If that sounds like an insult to your intelligence, you're not the customer and TextGen is the better tool.

Who should pick which

Pick text-generation-webui if you're on Windows, Linux or an Intel Mac; you want specific community models; you want an Anthropic-compatible API, or an API in front of a model you chose yourself; you want LoRA fine-tuning, image generation or voice; you want to serve other devices over your LAN; or open source is a requirement rather than a preference.

Pick Outlier if you're on an Apple Silicon Mac and you want something ready in one download, with no model shopping and no Terminal.

Running both is reasonable. One's a workbench, the other's an appliance.