TUFF

The local AI app built for the Mac. Runs models far bigger than your Mac's memory.

Download · Website · Benchmarks · Contribute · Discussions

Watch the film (66 s)

TUFF runs language models on Apple Silicon Macs. The app is SwiftUI and the engine is my own, written in Swift and Metal. It's open source, and nothing you type leaves your Mac unless you turn on web search.

Qwen3.8 Flash Next is a 111 GB model, and TUFF runs it on a 16 GB MacBook Air. TUFF keeps a mixture-of-experts model's experts on the SSD and reads only the few each token needs. GPT-OSS 120B and MiniMax M2.7 run the same way. On that Air, Flash Next writes about 1.7 tokens a second; smaller models are much faster. The benchmarks show how each model runs on real Macs.

Install

You need an Apple Silicon Mac on macOS 15 or newer.

1. Download the ZIP from the latest release, unzip it and move TUFF.app to Applications.
2. Open TUFF, go to Models and install one. Gemma 4 E2B is a quick first download.

TUFF isn't notarized, so the first time, Control-click the app and choose Open. Or use Homebrew, which also adds the tuff command:

brew install --cask tuff-app/tap/tuff

If you installed with Homebrew before TUFF moved to tuff-app/homebrew-tap, its README has the three commands that move your install over.

TUFF updates itself. If an update stops it from opening, see recovery.

What's in it

• Chat with reasoning, Markdown, math, and image and file attachments. Going back to a recent chat continues from where the model left off instead of reading the whole conversation again.
• Web and Files search. Off until you turn them on. The model can search the web or folders you pick, and cites what it used. How search works.
• Benchmarks. Measure any installed model on your Mac and share the result on the leaderboard. How benchmarks work.
• Background API. A local OpenAI-compatible server for agents and other apps. It loads whichever installed model a request names. Server and API.

Model	Download	Minimum memory	Images
Gemma 4 E2B Q4	2.6 GB	8 GB	Optional
Gemma 4 E4B Q4	4.2 GB	8 GB	Optional
Gemma 4 12B QAT	11.0 GB	16 GB	Optional
Gemma 4 26B-A4B Q4	14.3 GB	8 GB	Optional
Qwen 3.6 35B-A3B Q4	19.6 GB	8 GB	Optional
GPT-OSS 20B Q4	13.8 GB	16 GB	No
GPT-OSS 120B Q4	65.3 GB	16 GB	No
MiniMax M2.7 Q4	128.7 GB	16 GB, M2 or newer	No
Qwen 3.8 Flash Next Q4	110.9 GB	16 GB, M2 or newer	Optional
Ornith 1.5 9B Q4	5.1 GB	8 GB	Optional, 16 GB
Ornith 1.5 35B-A3B Q4	19.6 GB	8 GB	Optional, 16 GB
Ornith 1.5 122B Q4	68.7 GB	16 GB	Optional
Ornith 1.5 397B Q4	239 GB	16 GB	Optional
Qwen3.8 27B Q4	16.2 GB	8 GB	Optional, 16 GB
Gemma 4 26B-A4B Heretic Q4	15.0 GB	8 GB	Optional
Recursive Flint 40B A2.5B OptiQ	25.6 GB	16 GB	No
Bonsai 2 2B Q2	1.2 GB	8 GB	No
Bonsai 2 27B Q2	8.6 GB	16 GB	No

Models are grouped under MoE, Dense and Legacy. Each section can be collapsed. MiniMax and both GPT-OSS models stay available under Legacy.

The app downloads an inference engine when you install its first model. Models with compatible architectures share that engine. Auto chooses context, expert cache and resident layers together. A dense model that does not fit can stream layers from disk, but that can be slow. These are calculated choices; TUFF learns from comparable runs in the app and avoids configurations that caused paging. Memory and engine details.

Gemma Heretic is an optional abliterated derivative, converted by mlx-community from coder3101's checkpoint. Its routed experts and attention weights are Q4; shared MLPs and routers remain 8-bit. Its image pack comes from the same checkpoint. Reduced refusals do not mean better accuracy. The ordinary Gemma model remains available.

I test releases on a 16 GB M2 MacBook Air. Numbers for other Macs come from people's benchmarks. Images need an M2 or newer.

How TUFF compares

	TUFF	LM Studio	Ollama	Colibrì	TurboFieldfare
Built only for Apple Silicon Macs	✅	❌	❌	❌	✅
Desktop chat app	✅	✅	✅	◐	✅
Built-in web search in chat	✅	❌	◐	❌	❌
Search your own folders from chat	✅	◐	❌	❌	❌
Local OpenAI-compatible server	✅	✅	✅	✅	◐
Command-line tools	✅	✅	✅	✅	✅
Image input	✅	✅	✅	✅	✅
Dense models	✅	✅	✅	✅	❌
Own engine, not MLX or llama.cpp	✅	❌	❌	✅	✅
Streams MoE experts from disk	✅	❌	❌	✅	✅
Open-source desktop app	✅	❌	✅	✅	✅
Windows and Linux	❌	✅	✅	✅	❌
Models	18	GGUF / MLX catalog	Model library	13 engines, mostly MoE	Gemma 26B

✅ yes · ◐ partly · ❌ no. Checked against each project's docs on October 8, 2026.

• Web search: TUFF uses DuckDuckGo with no account, or Brave or Tavily with your key. Ollama's web search is a hosted API that needs an Ollama account. LM Studio has none built in, but you can add it with MCP servers.
• Folders: LM Studio can chat with documents you attach, which covers some of the same ground.
• Streaming means a fixed-size expert cache, not OS paging or CPU offload. LM Studio and Ollama have much bigger model libraries.
• Colibrì also streams experts and runs Flash Next, on more platforms. TurboFieldfare is where TUFF started.

This compares features only. I haven't benchmarked these against each other.

Command line

The app includes the tuff command:

tuff prompt "Explain mixture-of-experts models in two sentences."
tuff bench --models gemma4,qwen36 --share
tuff serve --port 8080

tuff serve runs the same server as the Background API. It only listens on your Mac and has no authentication. OMP setup is a tested coding-agent config.

Build from source

You need Xcode with Swift 6.2 or newer (I build with Xcode 27).

git clone https://github.com/tuff-app/TUFF.git
cd TUFF
bash Scripts/package_engines.sh dist/local-engines dist/local-private/signing-key 1.0.0
export TUFF_ENGINE_CATALOG_PATH="$PWD/dist/local-engines/engine-catalog.json"
swift build -c release
.build/release/TUFF

Clone builds keep models and chats in scratch/, so they don't touch an installed copy. Scripts/test.sh runs the tests; none need a model. Local 9.0 builds explains the downloaded family engines, isolated app packaging and local qualification. How TUFF works explains the engine and maps the code, and CONTRIBUTING.md covers building, testing and the rules for this repository.

Credits

TUFF began as a fork of drumih/turbo-fieldfare by Andrey Mikhaylov, which built the original Gemma runtime and expert streaming. Other credits are in THIRD_PARTY_NOTICES.md.

Contributing

TUFF is one project split across a few repositories:

Repository	What's in it
TUFF	The Mac app, the tuff command and the inference engines
sandbox	The Linux VM that runs web and file search in isolation
homebrew-tap	The Homebrew cask
tuff-app.github.io	The website and benchmark leaderboard

They all follow the same contributing guide, maintainer roles and code of conduct. Questions and ideas go in Discussions.

Security

Please don't open a public issue for a security problem. Report it privately through GitHub's vulnerability reporting on the repository it affects, or on TUFF if you're not sure which. The security policy has the details.

License

Apache 2.0. Model weights keep their own terms.