ResearchExplainerSep 30, 20267 min read

The Anatomy of Private AI, Part 1: What a Model Actually Is

A model is a folder of numbers you can download. This page lets you open one, read what’s inside, and switch the Wi-Fi off. Part 1 of a three-part series on why self-hosted open AI is private.

Andrew Liu
On this page

Say "open-source AI" to a room of partners or portfolio managers, and someone hears "client data ends up in public." The confusion is understandable: in software, open source means the code anyone can read, and people reasonably assume the openness points at their information too. It points the other way. Open means you can see and hold the model. Your data is exactly as private as wherever you choose to run it.

This is the first of three short explainers making that concrete. Today: what a model physically is, with one on this page you can open and look inside. Part 2 follows the question that actually matters, who can read what you type. Part 3 walks the security case for owning the whole perimeter.

The download
A model is a folder. The weights are the files inside it.
Everything an AI model is fits in this folder. Nothing in it runs until you run it, and it keeps working with the Wi-Fi off.
The download · click to open
Rotate
A 30-billion-parameter model · about 60 GB as a folder. One folder, nothing installed. Click it to open it.
Source: Folder and file sizes illustrate a ~30B-parameter model at 2 bytes per weight (about 60 GB); file count and names simplified from a real release layout.
Figure 1. A model, as it actually arrives: one folder you download. Open it, hover the files, look inside one, then switch the Wi-Fi off. Sizes are for a 30-billion-parameter model such as Muse Glimmer.

What's in the box

That folder is the whole product. Every modern AI model is two things. The first is the architecture: code describing how to do the arithmetic, and not much of it; the core of a transformer fits in a few hundred lines. The second is the weights: an enormous list of numbers that the training process tuned, saved to disk as a file. The architecture is the recipe; the weights are the cake.

The file is big but it is just a file: smaller ones run on a good laptop, and the largest need a server. Meta's Muse Glimmer has about 30 billion weights;2 at two bytes per number that's roughly 60 GB, a large download and a small hard drive. Kimi's K3, the strongest open model you can download,5 has 2.8 trillion;1 the same arithmetic puts it near 5.6 TB. Nothing about either file runs, listens, or transmits. A model at rest is as inert as a spreadsheet.

Claims like that are easy to write and easy to doubt, which is what the folder above is for. Open it, and the biggest things inside are files that hold nothing but numbers. Look inside one. Then switch the Wi-Fi off and ask it something: the answer comes from arithmetic running beside those numbers, in the room, and the folder is exactly as it was.

So what is a weight, then

A weight is one learned number: a tiny statement of preference, like "after the letter t, lean toward h." Training is the process of nudging billions of those preferences until the file, taken together, can continue text usefully. In a model you'd actually deploy, the preferences are over concepts rather than letter pairs, and there are billions of them, but that is a difference of scale, not of kind. A model is the frozen result of training: a file of numbers that arithmetic runs through.

Frozen matters. The weights were set during training, before you ever downloaded the file, and answering a question doesn't change them; the readout above said it plainly: the folder is unchanged, and it still answers. What a model "knows" about your business after answering a thousand of your questions is precisely what it knew before: nothing. Where your questions travel on their way to the arithmetic is a separate issue, and it's the entire subject of Part 2.

"Open" comes in four strengths

Once you know the model is a file, "open" stops being a vibe and becomes a question with a checkable answer: which artifacts do they hand you, and under what license?

The spectrum
Four strengths of open

The file is published. You can download, inspect, and self-host it. The license decides what you may build on it, so read it.

You get
  • The weights file
  • Self-hosting rights per license
You don’t
  • Training code or data
  • Always-permissive terms
For example
  • Kimi K3 — bespoke license with revenue and attribution clauses
  • Muse Glimmer 30B — plain Apache 2.0, no usage caps
Source: Kimi K3 license per Moonshot (Jul 2026); Muse Glimmer per Meta (Aug 2026); Mistral Large 3 per Mistral (Dec 2025); OLMo 3 per Ai2.
Figure 2. The spectrum, with current examples at each stop. Licenses are the fine print of openness: two models can both be 'open-weight' and carry very different terms.

The practical reading for a buyer: anything from "open-weight" rightward can live entirely inside your walls. The license column is where diligence happens, and the spread is real, K3's bespoke terms ask for a negotiation above certain revenue, while Muse and Mistral ship under plain Apache 2.0 with no caps.123 Ai2's OLMo goes furthest and publishes the recipe itself.4 The practical habit: treat a model license like any other contract, and have it checked like one.

What this buys you

Here's the asymmetry that the rest of this series builds on. With a service like ChatGPT, your text has to travel to the provider's computers, because that's where the AI is. With a model you download, the AI comes to your files instead. A file on your server, in your rack or your tenancy, has no channel home; it can't phone anyone because there is no phone in it, which is what the Wi-Fi switch above was there to show.

Honest limits, before the next installment oversells itself: the weights file is the easy part. Running it well takes a serving stack, patches, evaluations, and monitoring, and that operational layer, not the model, is where real security work lives. That's Part 3. And openness of weights isn't openness of recipe; only the far end of the spectrum shows you the training data. What the file guarantees is narrower and, for privacy, decisive: the model is an object you possess, not a counterparty you talk to. For a Canadian firm, that distinction is the shape of the confidentiality duty itself; the Law Society of Ontario's AI guidance is the document to read beside this series, and the question it asks is exactly the one above: where does client information travel?6

Sources

  1. Tom's Hardware, "Moonshot AI releases weights for Kimi-K3" (July 2026): https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-ai-releases-weights-for-kimi-k3-firing-a-shot-across-the-bow-of-openai-and-anthropic-open-weight-model-performs-almost-as-well-as-frontier-models-while-being-2-3x-easier-to-run
  2. Meta AI Research, "Introducing Muse Glimmer" (August 2026): https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
  3. The Register, "French AI shop Mistral rolls out full suite of Apache-licensed models" (December 2025): https://www.theregister.com/2025/12/02/mistral_3/
  4. Ai2, OLMo 3 release notes: https://allenai.org/blog/olmo3
  5. Interconnects, "Kimi K3: the open-weights escalation" (2026): https://www.interconnects.ai/p/kimi-k3-the-open-weights-escalation
  6. Law Society of Ontario, "Generative AI: Your professional obligations" (April 2024): https://lawsocietyontario-dwd0dscmayfwh7bj.a01.azurefd.net/media/lso/media/lawyers/practice-supports-resources/generative-ai-your-professional-obligations.pdf

Nothing to book here. Parts 2 and 3 land over the next two weeks, and after that we're publishing the proof instead of the argument: a recorded demo of the full Northwood platform running completely cut off from the internet, Wi-Fi off, answering from a model in the room. The file is the easy part; we build and run the rest, the access controls, the record of every question and answer, the upkeep, inside your environment. If you'd rather talk it through sooner, we're easy to reach.

Keep reading

More notes from the field.

All research