The Model Is Disposable. Your Notes Are Not.

On August 25, 2026, Apple introduced a desktop that can hold a frontier-scale AI model entirely in memory — up to 512GB of it ¹. The headline is the hardware. The lesson is quieter: once a machine can run any model, the model is the disposable part, and the plain-text notes you feed it are the only thing built to last.

Apple's own copy is blunt about the capability. Its M5 Ultra chip, inside the new Mac Studio, lets users "run huge LLMs with hundreds of billions of parameters entirely on device" ², with "1.2TB/s of unified memory bandwidth that is 50 percent higher than M3 Ultra" ². The 512GB ceiling itself is not the new part — the M3 Ultra Mac Studio already offered it in March 2025 ³; the jump is in bandwidth and compute. A model that used to need a data center now sits under a desk, and it keeps getting faster there. That inversion matters less for what it lets the model do than for what it exposes about everything around the model.

Hardware gets replaced. Weights get re-quantized, fine-tuned, deprecated, and swapped for next quarter's release. The document you wrote is the one component in the stack with no version number and no end-of-life date. Steph Ango named the general shape of this before the hardware caught up: "Apps are ephemeral, but your files have a chance to last" ⁴. Trade "apps" for "models" and the sentence still holds.

We keep score on the model. That is the wrong scoreboard.

The industry measures progress by the model: parameter counts, benchmark ranks, which lab leads this month. It is a reasonable instinct, because the model is the part that feels intelligent. But the model is also the part that changes fastest, and building your working life around the smartest current model means rebuilding it every time a smarter one ships.

Simon Willison, who has spent years running these models on ordinary hardware, defines the open-weight ones plainly: they are "models that have been shared publicly by their creators and that can, in principle, be downloaded and run with local hardware" ⁵. That "in principle" is now "in practice." But the shared, downloadable model is also endlessly forkable and replaceable. There is always a better checkpoint next week.

Benchmarks make this vivid. The leaderboard that crowns a model in the spring is reshuffled by autumn, and the reshuffling is the healthy sign of a field moving fast. A fast-moving field is a wonderful thing to borrow from and a poor thing to build on. You want to ride the improvement without wiring your permanent record into whatever is winning this quarter.

The model is a lens. You do not build a library around a lens.

Models are the fastest-churning layer in the stack

Treating the current model as your foundation fails for one measurable reason: models turn over faster than almost anything else you touch. They are abundant, cheap to produce, and constantly superseded. What feels permanent this quarter is a rounding error in a catalog growing by the hundreds of thousands.

The scale is easy to underestimate. Hugging Face, the main hub for open models, passed three million public models in August 2026; the milestone was made official on August 18, and by August 22 the Hub counter stood at "3,012,377 public models," with roughly 750,000 of them added in just the first 7.7 months of 2026 ⁶. Weights get re-quantized to fit new memory budgets, retuned, renamed, and retired. Hosted models are sunset on vendor schedules. None of that churn is a problem, right up until you have stored something you care about inside it.

There is a sharper version of the problem. The model you fine-tuned on your own material last year is often impossible to reproduce exactly today: the base checkpoint moved, the quantization changed, the hosted endpoint was retired.

Software has always aged this way. The formats that survived the last four decades of computing were the open, documented ones, plain text among them, not the applications that read them. WordStar is gone; the letters typed in it and saved as text still open. A smarter model does not change that pattern. It only raises the stakes, because we are now tempted to pour more of our thinking into the layer that ages fastest.

A model you cannot export from is not a tool. It is a landlord.

The durable layer is the file you own

What survives every upgrade is the source of truth, and the most durable source of truth is a plain-text file on a device you control. It has no runtime, no login, no subscription to lapse. Any model, this year's or 2030's, can read it. That is the whole trick: keep the intelligence swappable and the data still.

This is the local-first argument, made years before frontier models fit on a desk. In 2019, Martin Kleppmann and colleagues at Ink & Switch put the principle in a paper's subtitle: "You own your data, in spite of the cloud" ⁷. Their point was ownership and longevity: the copy on your device should be the primary copy. Local AI finishes the picture that paper started. Now the computation can be local too, so nothing about the intelligent part requires shipping your notes anywhere. The model visits your files; your files never move.

The asymmetry is what makes the habit easy to keep. A file costs nothing to hold in a form every tool understands, and it quietly gathers value as models improve around it. Each new model that can read plain text makes the notes you already own more useful, with no migration, no conversion, and no re-upload. You are not betting on which model wins. You are making sure that whichever one does, it has to meet your notes on your terms rather than the other way around.

The practical version of this — pointing a model at notes that never leave your machine — is one we have written about before, in why local AI is becoming the default.

What to do while the models keep changing

You do not need the new desktop to act on any of this. You need one habit: keep the thing you actually care about — your notes, your writing, your records — in an open format you can read without a specific app, then let models come and go against it. Five moves make that concrete.

1. Write the source of truth as plain text. Markdown is enough; it opens in anything and will still open in a decade.
2. Keep the file on a device you control, not locked inside a service you rent.
3. Treat the model as swappable. Point whatever is best this month at your notes, and expect to point a different one next year.
4. Never store the only copy of an idea inside a model's memory or a vendor's cloud. Those are the parts that expire.
5. If you do send text to a hosted model, decide it on purpose. A local model keeps the text on your machine; a cloud one does not.

Frequently asked questions

Can you really run a large language model locally on a Mac? Yes. Apple's 2026 Mac Studio with the M5 Ultra offers up to 512GB of unified memory, and Apple says the M5 Ultra lets users "run huge LLMs with hundreds of billions of parameters entirely on device" ². Smaller models have run on ordinary laptops for years; the new hardware extends that reach to frontier-scale ones.

Does local AI replace your notes app? No, and that is the point. A model reads and transforms text; it is not where the text should live. Your notes are the durable source of truth. The model is a swappable lens you point at them — useful, but disposable, and never the place to keep the only copy of anything.

Why keep notes in plain text instead of inside an AI tool? Because plain text outlasts the tool. A Markdown file has no runtime, login, or subscription, and any model, current or future, can read it. Weights get re-quantized and deprecated; a format you own does not. Ownership and longevity are the local-first argument Ink & Switch made back in 2019 ⁷.

How often do AI models actually get replaced? Constantly. The Hugging Face Hub passed three million public models in August 2026 and stood at 3,012,377 by August 22, with roughly 750,000 added in the first 7.7 months of 2026 alone ⁶. Open weights are re-quantized and retuned continuously, and hosted models are retired on vendor schedules. The current "best" model is always temporary.

Is running AI locally more private? It can be, on the paths that stay local. A model running on your own machine processes your text without uploading it. But if you opt into a cloud model, the text you send leaves your device to that provider. Privacy depends on the specific path, so choose each one deliberately.

What is the durable part of an AI setup? The data. Hardware is replaced, weights churn, and apps come and go, but the files you wrote carry no version number and no end-of-life date. As Steph Ango puts it, "Apps are ephemeral, but your files have a chance to last" ⁴. Keep the intelligence swappable and the source of truth still.

The desk-sized frontier model is a genuine milestone, and it will be superseded by the time you finish reading its spec sheet. That is exactly why it clarifies things: when the smartest part of your setup is the part most certain to be replaced, the only sane place to keep what matters is a file you own.

MNMNOTE keeps your notes as plain Markdown on your own device, readable with or without it — the still center while the models keep changing: mnmnote.com.