Innhush

A private AI in a box for small hotels. It answers guests in their language from your own documents and shows where every fact comes from. Runs fully offline on your Mac — guest data never leaves the building.

Download for Mac One-line install

Mac with Apple silicon (M1 or newer), macOS 14+, 8 GB memory or more. Free models (~3 GB) download once.

What it does

Your documents

Drop in house rules, prices, contracts, FAQ — txt, md, pdf, docx. Innhush reads them and searches them in a split second.

Replies in the guest’s language

English, Русский, Bahasa, 中文, Deutsch and more. Facts get a mark like [1] and a “Where it comes from” list.

Says “not in the documents”

If the answer is not written down, it does not invent one: “I will check with the manager.” A second pass removes any sentence the documents do not support.

Hotel templates

Check-in and check-out, breakfast, transfer, house rules, complaint — one click, then edit.

Menu-bar icon

A box with a coloured dot: green ready, blue writing, yellow starting, red error. Answer the question in the clipboard with ⌘R.

Works offline

The model runs on this Mac (llama.cpp, Apple Metal). Network: 127.0.0.1 only. No account, no cloud, no telemetry.

How it looks

Measured on a real Mac

Questions from guests	20 in 6 languages: 20 of 20 correct, 3 of 3 “not in the documents”
Speed of a reply	51–81 tokens/s while writing (llama-bench: 87 tokens/s), 0.3–1.9 s per reply
Memory	about 3.8 GB for the reply model + 1.0 GB for search
Network	lsof and nettop during the test: only 127.0.0.1, 0 bytes outside

Mac with M5 Pro, 64 GB. Model: Qwen3 4B Instruct 2507 (Q4_K_M) + Qwen3 Embedding 0.6B. Both are free and open (Apache 2.0).

Install in one step

Paste into Terminal:

curl -fsSL https://innhush.sudarev.app/install.sh | bash

Or download Innhush.zip, move Innhush to Applications and open it (signed with Apple Developer ID and notarized — opens without warnings). Then press Install in one step — the free models download once, and from then on it works without internet.

Which Mac to choose

Memory	Model	Good for
8 GB	Qwen3 1.7B	A guesthouse, short replies; slower on long documents
16 GB	Qwen3 4B (recommended)	A small hotel, one front desk — the best value (e.g. Mac mini)
24–32 GB	Qwen3 4B with headroom	Many documents, other apps open at the same time
64 GB+	Larger open models	Several staff at once, long contracts, the best quality

Memory matters more than the processor: the model lives in memory. Choose at least 16 GB if you buy a Mac for Innhush.

Private by design

Documents, questions and replies are stored only in ~/Library/Application Support/Innhush on this Mac. The model server listens on 127.0.0.1 only. The only time Innhush goes online is the one-time model download from Hugging Face, which you start yourself. Reply history can be cleared from the menu.

Free now. A paid licence switch is built in for later and is off.