Products
Base Router, the always-on AI box we are designing
A preview of Base Router, a planned $400 private AI box built on a Jetson Orin Nano Super, with its speed targets, its limits and its consent design.
Base Router is the second product on offlinebase.com, and we are still designing it. The Base Router page (opens in a new tab) carries a plain status line: "Preview only. Base Router is not available to order." Everything below describes Base Router as our site shows it as of October 2026.
The planned price is $400, paid once, with no subscription. The tagline is "The full package, fine-tuned to you."
Base Router after Base Drive
Base Router is the step after Base Drive. Base Drive is a USB drive that sets up private AI on a laptop you already own. Base Router takes the same private AI and moves it onto dedicated hardware that stays on. Base Drive is coming soon too, at a planned $19.99, and neither product can be ordered yet.
We list Base Router as best for "a capable everyday assistant, document work over your own files, and field-specific setup." For the wider picture of the company, read What Offline Base is building.
A finished appliance
Base Router is planned to arrive with the software and AI already loaded. It is meant to join your network without a keyboard, a monitor or technical setup. The headline on the Base Router page puts the goal in one line: "One box. One promise: it just works."
What's inside
Base Router has four layers.
Layer | What it is |
|---|---|
Aluminum shell | The case, with no microphone and no camera |
Compute | A Jetson Orin Nano Super, a small NVIDIA computer that runs the AI |
Storage | Where the models and your documents live as ordinary files |
Quiet cooling | Active cooling with a small fan, steady through long sessions |
We left out the microphone and camera on purpose. As the page puts it, "an appliance that holds your private work shouldn't have either." The AI itself is a set of files on the storage layer, yours to inspect, and your documents stay on that same layer. The page adds: "Nothing inside listens, watches, or phones home."
Inside the Jetson Orin Nano Super covers the chip's specs, memory and price history.
You can get a feel for the board today. offlinebase.com/try (opens in a new tab) is a live chat with Liquid LFM2.5 1.2B running on a Jetson Orin Nano Super that we operate. That model is far smaller than the 8B and 9B models planned for Base Router, so it types faster than the targets below.
Power and speed targets
Base Router is planned to draw 7 to 25 watts. How that compares with a data center is the subject of The energy case for local AI.
Two models are planned, each with a target speed:
Model | Target speed | What it is for |
|---|---|---|
IBM Granite 4.1 8B | about 15 tokens/s | Tuned for documents and business tasks |
Qwen 3.5 9B | about 12 tokens/s | Strong reasoning, the heavier and slower pick |
Tokens per second is how fast a reply types out on screen. The 8B and 9B are model sizes in billions of parameters. Bigger models tend to give answers that read better, and they generate more slowly, which is why the larger Qwen model has the lower target.
Every one of these numbers is a target. The page calls its speed figures "targets and third-party benchmarks" and says they carry no guarantee. Real speed depends on the model, the length of the task and what else the machine is doing.
The speed preview on the page asks for a summary of a client memo ("Summarize this 2-page client memo into a few bullet points."). You pick a model and the answer types out at that model's target rate, so you can watch what 15 or 12 tokens a second feels like.
Limits we state up front
The Base Router page has a section headed "We'd rather under-promise." Here is what it covers.
Smaller than the biggest cloud AI. The AI in Base Router is built to be capable and useful for everyday work. The largest cloud systems will still do better on the hardest questions. For those, the plan is for Base Router to ask your permission before using outside help, which would always be optional. Outside help is not in any shipping build yet, and the consent section below covers how it would work.
A steady conversational pace. The targets are about 12 to 15 tokens a second, and very long answers and large documents take more time.
"Updates can ship on a USB stick." Each update is checked before it installs, and the previous version is kept so you can roll back. In the page's words: "Offices that stay offline can stay offline."
It runs warm. The small fan from the table above keeps it steady through long sessions.
Asking before anything leaves the box
Everything in this section is planned.
Our How it works (opens in a new tab) page states the rule: "Local by default. Cloud only with your explicit yes." The plan includes:
- A prompt before a single word leaves the box
- A written record of every approval
- A local-only mode that turns outside help off entirely
As designed, outside help stays off until you turn it on. The provider rules it would follow are in What zero data retention means, and where it stops.
Remote access is planned as an optional feature. It would start switched off, and you would have to choose to turn it on. The connection is designed to be encrypted end to end.
Our reasons for putting local first are in Why we make local the default for private AI. If you run a clinic, a law firm or a school, What on-premises AI can and cannot do for clinics, law firms and schools covers office use.
No launch date yet
Base Router has no launch date yet. Our terms (opens in a new tab) treat every specification, price, model list, speed figure and date on the site as a current target, so the $400 price, the power figures, the model list and the speed targets can all change before launch. The current design lives on the Base Router page (opens in a new tab). The Offline Base Letter (opens in a new tab) is our newsletter, and Base Router news will appear there.