Can AI run without the internet? Yes for the model itself. A few parts still need a connection.
Short answer. Yes. Once a model’s files are on a machine, tools such as Ollama, llama.cpp and Open WebUI generate answers on that machine without calling out to the internet. A connection is needed for the first download of a model (17 GB for Gemma 3 27B in Ollama’s library), for updates, and for anything that uses the web by nature: search, cloud models and online lookups. When you evaluate a product, ask what stops working when the network cable is pulled, because some products sold as local still contact the vendor for licences or features.
01
What works with no connection
Task
Works offline?
Why
Chatting with a model already on the machine
Yes
Ollama serves models from the machine and binds to 127.0.0.1 by default; llama.cpp is built for local inference
Answering from your own documents
Yes, if documents and index are stored locally
Open WebUI’s README describes local retrieval-augmented generation with built-in vector databases
Transcribing a meeting
Yes, with a local speech model
Whisper’s code and weights can be downloaded and run on your own machine
Reading a scanned page
Yes, with local OCR
Docling documents local execution for sensitive data and air-gapped environments; OCRmyPDF adds a text layer using Tesseract
Searching the web
No
It needs the internet by definition
Calling a cloud model
No
The request leaves the building; ChatGPT, for example, runs on OpenAI’s GPUs
Downloading or updating a model
No, unless files are copied across
Models are pulled from a library such as Ollama’s; Hugging Face’s offline mode reads only files already cached
02
How big the first download is
The one-off cost of going offline is the download. These sizes are from Ollama’s model library; the times are plain arithmetic at a steady 100 Mbps, not measurements, and a model can also be copied across on a drive.
Model in Ollama’s library
Download size
Time at a steady 100 Mbps
Gemma 3 4B
3.3 GB
about 4 minutes
Gemma 3 12B
8.1 GB
about 11 minutes
Gemma 3 27B
17 GB
about 23 minutes
Qwen3 32B
20 GB
about 27 minutes
03
Questions to ask a vendor
What happens if the connection is cut for a week?
Does the product check a licence online, and what happens if it cannot?
Which features call a cloud service, and can an administrator switch them off?
How do updates reach a site with no internet?
Is every cloud call labelled and logged, so you can see if anything left?
04
How ANVE handles it
ANVE’s text models run on the box. Cloud models are optional: an administrator has to add a provider, and their answers are labelled; a policy that blocks confidential data from cloud models is not on ANVE yet. Every model call is logged with the user, app, model and whether it ran locally or in the cloud. Two parts reach outside today: the Scholar-Atelier research maker looks sources up online only when asked, and podcast narration uses a cloud voice while local narration is not on ANVE yet.
On updates and licences: ANVEAI installs updates on site today, signed update packages applied offline by USB are not on ANVE yet, and ANVE does not update itself over the internet. If ANVEAI shuts down the box keeps working, every order includes a perpetual licence for the software delivered with the unit, and there is no remote kill switch. The sealed unit, which is in design, adds two physical switches with lights: one cuts the network and one cuts the microphones. ANVE is pre-release, and the software runs on a development machine. The details are on the security page.
05
Questions people ask
Does ChatGPT work offline?
No. ChatGPT runs in OpenAI’s cloud, where model inference runs on OpenAI’s GPUs, so every prompt goes to OpenAI’s systems. A local model avoids that, at the cost of smaller models.
Do I need the internet to update a local model?
To fetch a new model, yes, or you copy the files from another machine. Hugging Face’s tools have an offline mode that makes them read only files already cached on the machine.
Can an air-gapped setup run AI?
Yes. Docling documents local execution for air-gapped environments and Open WebUI is built to run offline. Updates then have to arrive by another route, such as files copied across under your own change control.
Is a local model slower without a connection?
No. Speed depends on the hardware and the model size, not on the connection.
06
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
