The calm, knowledgeable friend in the room — even when everything else is offline.
Sage in a Box is Sage, running entirely on your own computer. No cloud, no account, nothing sent anywhere. The download is small; the first time you open it, Sage fetches its engine, a model and a medical encyclopedia and sets itself up. After that it needs no internet at all.
If the internet disappears tomorrow, your Sage keeps working — that's the whole point. When you are online it can quietly learn from the live Sage and keep what it learned for the days you aren't.
Windows 64-bit · free · no account, no installer. Not sure which? Read the next section — and it isn't a decision you can get wrong, because Lite becomes Beefy later from its own Settings page without downloading everything again.
66downloads
1box linked to live Sage
Optional: linking lets your box learn from the live Sage while you're online.
Build your box
The download is the same small zip whichever you pick. What differs is what Sage fetches the first time it opens - a model or five, an encyclopedia or the whole of Wikipedia - and you choose that here, with real sizes, then get a short code. Sage asks for the code once. Everything you skip stays visible inside Sage with its size and a Download button, so nothing here is a decision you can get wrong.
A starting set, not a commitment. Tick and untick anything below; the box downloads exactly this list the first time it opens, the essentials first, the rest in the background while you use it.
The model that does the thinking
Pick one to start with; more than one is fine on a big machine.
Eyes
A second model that reads photos. Needs its own room in memory.
Search
Comes with every Sage. Nothing to choose.
The library
Whole encyclopedias, from Kiwix, straight onto your machine. The big ones take days on a slow connection and never hold anything else up.
Your box will download5.5 GBthe first time it opens, engine included.
Wants about 2.2 GB of memory free for the smallest model you picked.
Then download. The code goes in on the first run; the zip is the same for everyone.
Lite or Beefy?
The difference is how much computer you are giving Sage, not how good the answers are. An AI model has to fit in your computer's memory to run at all, so Sage ships in two shapes: one that is careful about memory, and one that spends it. Both are the same Sage, and both work with the internet unplugged.
Sage Lite
For the old laptop, the 8 GB desktop, the machine in the garage.
Chat and the offline library, tuned to stay out of the way. A small model, one thing in memory at a time, and a memory budget you can see and change. This is the right choice for a modest computer, not a stripped-down one — an old machine that keeps a family answered when the power is out is the whole point.
Chat, entirely offline
The offline library and search
Plain-language settings and a memory panel
SageBeefy
For the desktop with room to spare.
Everything in Lite, with the workshop switched on as it lands: your files, USB devices, cameras and boards, and Labs that keep running on their own. Same clear settings, more switches — bigger models, more than one loaded at once, and as many devices as your machine can carry.
Everything Lite has
Bigger models, more than one at a time
Files, devices, cameras and Labs as they land
How to tell which one this computer wants. Open Task Manager, look at how much memory the machine has, and pick Lite if the answer is 8 GB or less. If it is 16 GB or more, take Beefy. If you take the wrong one, switch inside Sage — nothing you have downloaded is thrown away.
Three things it's for
🩹
Keep my family safe
A child with a high fever at 2 a.m. A deep cut. Someone who collapsed. Sage walks you through what to do right now, what to watch for, and how urgent it really is — with weight-based doses and step-by-step checklists.
“My 4-year-old has a 103° fever and a rash”
“How do I stop heavy bleeding?”
“Make oral rehydration solution”
🏕️
Keep my homestead running
Safe water, food that won’t spoil, staying warm without power, sanitation when the utilities quit, keeping a small solar or battery system alive. Practical know-how that works when the grid doesn’t.
“Is this creek water safe to drink?”
“Heat one room with no electricity”
“Size a 12V solar setup for a fridge”
🔧
Build and learn
Ask how a thing works and read the article behind the answer. Sage carries whole offline encyclopedias — medicine today, repair guides, appropriate technology and all of Wikipedia if you want them.
“Explain how a well pump works”
“What does a soil moisture sensor actually measure?”
“Search WikiMed for “heat exhaustion””
What you get
The download is small on purpose. Everything below either ships in it or is fetched once, from inside Sage, while you watch a progress bar.
💬
Chat that runs on your machine
Ask anything. The answer is generated on your own computer by your own model. Nothing is sent anywhere, and no account is involved.
🩺
WikiMed, the first thing Sage fetches
Every medical article on Wikipedia, downloaded once during setup and indexed on your machine. After that it works with the internet unplugged.
📚
A library that swallows encyclopedias
Point Sage at a folder — an external drive, a USB stick — and it indexes every .zim archive it finds. Search them, read them, ask Sage about the page you are on.
🧠
A brain sized to your computer
Sage downloads a small, fast model on first run, and a smaller one is a click away if this machine wants it. You can see exactly how much memory it is using.
🎚️
Settings written for people
Every control says what it does and when you would change it. Memory, model, context length — no jargon, and a straight answer about what would happen if you turn it up.
🔄
It keeps itself current
Sage checks for a newer version when it starts and installs it in place. Your conversations, settings and library are yours and stay put.
🔗
Sage Link
Optional, and never required for anything. When you are online, your box can ask the live Sage the same question you just asked it, keep the answer, and carry it into the days you are offline.
🔒
Yours, on your terms
One folder. Copy it to a USB drive and it is the same Sage on another machine — same conversations, same library, same settings.
🧰
The workshop, arriving
Files, USB devices, cameras, boards and Labs are being built now and land in SageBeefy as they are ready. An installed box updates itself into them.
Up and running in four steps
📦
1. Unzip it
The download is small — a few hundred megabytes, not a night of waiting. Right-click the zip → “Extract All”. Inside: Start Sage, and the app and updater folders. That is the whole install; put it on a USB drive if you like.
🖱️
2. Double-click “Start Sage”
The only file you ever open. Sage starts on your machine and your browser opens on it. Nothing is installed into Windows, and it never asks for admin.
⬇️
3. Let it set itself up
On the first run Sage fetches the AI engine, one small model and WikiMed — a few gigabytes, once, with a progress bar you can watch. Do this while you still have internet. Every launch after that takes seconds and needs nothing.
💬
4. Meet Sage
Ask it something. Then unplug the internet and ask it something else — that is the part worth trying on the first day.
Choose your brain
The model is the part that does the thinking, and a bigger one is slower and hungrier for memory rather than simply better. Sage downloads a small, fast one on first run and offers a smaller one for machines that need it, in Settings → Model. Switch any time; the download happens once and stays on your disk.
qwen2.5:3b
~1.8 GB
Downloaded on first run
Small, quick and surprisingly capable. Answers in a few seconds on an ordinary CPU with no graphics card at all, and leaves enough memory free that your browser stays usable while it thinks.
The default, and the right one for most computers.
qwen2.5:1.5b
~940 MB
One click in Settings
Half the size and about twice as fast, and it shows: shorter answers, less depth. Still genuinely useful for "what do I do right now" questions.
For a machine where the default feels slow. Better a fast answer than none.
llama3.1:8b
~4.6 GB
With SageBeefy
Careful, factual, strong at step-by-step reasoning, and the most reliable of the bunch at using Sage's tools. It will occasionally hedge on medical or risky topics even with Sage's instructions.
The best answers, on a computer with room to spare. Best balance of accuracy, good judgment and tool use for family, medical and homestead questions.
qwen2.5:7b
~4.4 GB
One click in Settings
Excellent general knowledge, strong at math, tables and long structured answers; follows Sage's instructions very closely and rarely hedges. Uses tools well.
Great second choice. Pick this if Llama feels too cautious or you want more detailed technical answers.
mistral:7b
~3.8 GB
One click in Settings
Fast, direct and plain-spoken with very few built-in refusals. Slightly less knowledgeable than Llama or Qwen on deep medical detail.
Best for older or slower computers among the big ones, and for people who want blunt, no-nonsense answers.
gemma2:9b
~5.0 GB
One click in Settings
The strongest reasoning in this size class and very good at explaining things clearly. Needs a little more memory than the others. Does not support tools, so Sage cannot act while it is selected.
Pick this on a well-specced machine when you want the smartest offline answers and do not need Sage to act.
qwen2.5-coder:7b
~4.4 GB
One click in Settings
Tuned for writing and debugging code and firmware. Weaker at general conversation.
Only for the workshop: Arduino sketches, scripts, wiring logic.
qwen3.5:9b
~6.1 GB
One click in Settings
A newer generation than Llama 3.1 in about the same room. Uses tools, and it can look at pictures itself, so on a machine that runs it there is no swap to a separate pair of eyes.
The all-rounder for a 16 GB desktop: chat, files, boards and eyes in one model.
gemma4:12b
~7.1 GB
One click in Settings
Google’s open model: strong plain-language answers, tools, and pictures. A touch bigger than Qwen 3.5 9B.
A second opinion at the same tier as Qwen 3.5 9B; try both and keep the one whose voice you like.
gpt-oss:20b
~13 GB
One click in Settings
OpenAI’s open-weight model. Reasons before it answers, so it is slower on a CPU and better at the hard ones. Uses tools; no pictures.
A 32 GB desktop that wants the strongest answers and can wait for them.
qwen3.5:27b
~16 GB
One click in Settings
The big Qwen: tools, pictures, and answers close to the online services on most everyday questions. Needs a real desktop and a fast CPU or a GPU.
32 GB or more, and patience on a CPU. This is the decked-out box.
llama3.3:70b
~40 GB
One click in Settings
Meta’s largest open model at a size a home machine can hold. Uses tools; no pictures. Slow without a big GPU.
64 GB and a GPU, or a lot of patience. As far as offline goes today.
Give Sage a bigger library
WikiMed arrives with Sage. Beyond it, Sage can hold entire offline encyclopedias: download free .zim knowledge packs from Kiwix onto any drive, then point Sage's Library page at the folder. It indexes every article and answers from it with the internet unplugged. Do this while you still have internet — the packs are big, and an external drive is a fine place for them.
Appropedia (~555 MB)
Solar, water purification, gardening, off-grid building. Appropriate technology for the homestead, and the best next download.
Open folder ↗ and grab the newest appropedia_en_all_maxi_*.zim
iFixit (~3.3 GB)
Repair guides for thousands of devices: phones, laptops, appliances, tools.
Open folder ↗ and grab the newest ifixit_en_all_*.zim
Wikipedia (no pictures) (~49 GB)
All of English Wikipedia as text. The whole encyclopedia, on your machine.
Open folder ↗ and grab the newest wikipedia_en_all_nopic_*.zim
Wikipedia (with pictures) (~115 GB)
The whole encyclopedia with its pictures. Yes, really. Put it on a big drive.
Open folder ↗ and grab the newest wikipedia_en_all_maxi_*.zim
Project Gutenberg (~206 GB)
Sixty thousand free books, with their covers. A library for the long haul, and by far the biggest thing here.
Open folder ↗ and grab the newest gutenberg_en_all_*.zim
Sage and her big sister
Sage (online, here on the site)
Large model, always current, has the internet
Acts on your account: creates projects, files journal entries, attaches photos
Shares with family members you choose
Sage in a Box (offline, yours)
Runs on your hardware; works with the internet off
Same voice and purpose, sized to the computer it is on
Learns from online Sage only if you link it — and keeps what it learns
Linking the two is optional and never required for anything. When it is linked and online, your box can ask the live Sage the same question you just asked it and show you both answers, marked so you always know which is which. Keep one and it becomes part of what your box knows — including on the day the internet is gone. Unlinked, or off the grid, every part of Sage still works; you simply see one card instead of two.
What you'll need
A Windows 64-bit desktop or laptop. 8 GB of memory and no graphics card is enough for Sage Lite — that is a deliberate target, not a grudging minimum. 16 GB or more is where SageBeefy belongs.
About 10 GB of free disk space for Sage, its model and WikiMed. Encyclopedias and extra models are on top of that, and they can live on any drive you point Sage at.
An internet connection for the download and the one-time setup. After that, none — and that is the part worth testing on your first day.
The setup we recommend
Sage in a Box is one folder, and it runs from wherever that folder sits. So the best way to own it is not on a computer at all: it is on a drive, whole, and the computer is just whatever you plug the drive into.
🖥️
The computer
Any Windows desktop works, and an old 8 GB one in the garage is the machine this was built for. The one we would buy: a desktop with 32 GB of memory or more, a modern 8-core processor, an NVIDIA graphics card with 12 GB or more — Sage's engine uses it, and answers that take twenty seconds on a CPU take two — and plenty of USB ports, because the workshop is going to fill them with boards, cameras and a drive.
💾
The drive: LaCie Rugged SSD, 4 TB, USB-C
Put Sage on this and carry it. Four terabytes holds Wikipedia with its pictures, every book on Gutenberg, every model in the list and years of what your box has learned, with room to spare. It is an SSD, so a model loads in seconds rather than the half-minute a spinning drive takes, and it is rugged, so the garage, the truck and the trip to a family member's house are all fine. Plug it into any Windows computer and double-click Start Sage; it is the same Sage, linked as the same box, with everything it knows. Close Sage before you unplug it, and format it exFAT or NTFS, never FAT32.
Why offline?
Stay Active · Live Free is about resilience. Sage in a Box is that idea applied to knowledge: a self-reliant companion that lives on your hardware, knows how to keep a family safe and a homestead running, and grows with you. Connect it and it learns from the live Sage; cut the cord and it keeps everything it has learned.
Versioned, updated in place, and yours. Expect it to get better with every release.
One honest note: Windows may put a blue “unrecognised app” screen in front of Start Sage the first time, because it is new and not yet code-signed. Choose More info → Run anyway, or use Start Sage (backup).cmd in the same folder, which Windows leaves alone. Signing is on the way.