# Run the whole thing in your building
> Apple shipped a desktop that holds 512GB of model weights at 1.2TB/s. Paired with a local General Input deployment, the prompts, the weights, the credentials, and the execution never leave your network.
Published: 2026-08-27. Last updated: 2026-08-27.
On Tuesday, Apple introduced a new Mac Studio built around a chip called the M5 Ultra ([Apple](https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/)). The number Apple led with wasn't how fast it computes. It was how much memory it holds and how quickly it can read it: up to 512GB, moving at 1.2 terabytes per second. Apple's hardware chief called it "the ultimate desktop for on-device AI."
If you're not deep in this world, that sounds like a spec-sheet detail. It isn't. It's the sign of a change in what a powerful computer is even for.
## The number on the box used to be speed
For about thirty years, you bought a workstation for raw computing power. That made sense, because the work that justified an expensive machine was math-heavy: rendering video, running simulations, compiling code. More calculations per second meant the job finished sooner.
AI doesn't work that way. When a model writes an answer, it produces one word at a time, and for every single word it has to read its entire brain back out of memory. It barely does any math along the way. Almost all of the waiting is reading.
So two numbers decide everything. How much memory you have decides whether a model fits on your machine at all. How fast that memory can be read decides how quickly it answers. Raw computing power is now the third number down the list.
If you're comparing machines for AI, read the memory line first and the speed
line second. A slower chip with more memory can run models that a faster chip
simply cannot load.
## What 512GB actually gets you
Here's a concrete example. DeepSeek publishes a model called V4-Flash, and the file is 167GB. It's free to download and openly licensed.
That size is the whole story. It won't fit on a laptop, or a Mac mini, or even on a single one of the datacenter graphics cards that cost thirty thousand dollars. To run it on smaller hardware you have to compress it first, which means running a rough copy of the model instead of the real thing.
On a 256GB Mac Studio, it loads whole. No compression, no compromise, and no meter running while you use it.
That machine isn't cheap. Configured to the top of what Apple currently sells, it's $18,299, which is more than most people expect a computer without a screen to cost. But it's a thing you buy once, and nobody bills you per question after that.
## A model on your desk is only half the job
Running the model locally answers where the AI thinks. It doesn't answer what the AI is allowed to do.
Real work isn't answering questions in a chat window. It's pulling a record from your CRM, checking a calendar, drafting the follow-up, updating the invoice. The moment AI starts doing that kind of work, it needs access to your accounts, which means your passwords and keys are in play. And if the software coordinating all of that lives in someone else's cloud, then your data is leaving the building anyway. The model stayed home, but the work didn't.
That's the part we built General Input around, and it's why the local model story only recently became a complete one. General Input runs on your own infrastructure. The dashboard, the API, and your database are ordinary services you deploy where you like. When an agent writes and runs code, it does that inside a container on your machines rather than in our cluster. File storage speaks the standard S3 protocol, so a server in your own rack works the same as a cloud bucket.
Because the software is now on your network, pointing it at the Mac Studio is simple. You run a model server on the machine, add its address in settings, and it shows up in the model picker beside everything else. It's just another computer on the same network.
Put that together and the loop closes. The request comes from a server in your building, gets answered by a model in your building, runs code in your building, using credentials that never left your building.
## What still reaches out, and why
One piece doesn't run locally, and it's worth being precise about which.
General Input documents tens of thousands of API operations across hundreds of services, so your agents know how to actually talk to Salesforce or Stripe or NetSuite. That documentation changes constantly as vendors ship updates, and it's identical for every customer, because it describes public APIs. A local copy would be out of date within a week.
So a local install reads that documentation from us as it needs it. That request asks what the Salesforce bulk API expects in a request body. It doesn't carry your Salesforce data.
Ask any vendor promising on-premise AI the same question: what exactly crosses
the network, and what's in it? "Nothing ever leaves" is almost never true. The
useful answer is a specific one.
Your credentials get the same treatment they always have. They're encrypted at rest, injected only when a step needs them, and stripped out before anything reaches the model. Every action taken on your behalf is logged. That holds whether the model is a cloud service or a machine humming under the desk.
## The honest case against doing this today
Two things argue the other way.
Open models still trail the best commercial ones by roughly six to twelve months, and they're weaker on unusual problems that look nothing like their training data. A desktop runs a very good model. It doesn't run the best one, and for some work that gap will matter.
The other is timing. Memory is having its worst pricing year on record. A common 32GB memory kit went from around $120 in late 2025 to nearly $400 this August, and the Mac mini's starting price climbed from $599 to $899 in five months ([Tom's Hardware](https://www.tomshardware.com/desktops/mini-pcs/apple-price-hikes-continue-as-mac-mini-with-16gb-ram-and-256gb-is-now-usd899-1tb-storage-option-adds-usd500-to-entry-level-headless-system)). The capability is heading toward being included with the hardware. The component is getting more expensive on the way there. Both are true, and the second one is a real reason to wait.
## The takeaway
The expensive part of AI was never the model. Models get cheaper and better every few months, and the one you're loyal to today will be replaced.
What's actually hard to replace is everything around it: the connections to your systems, the credentials, the accumulated knowledge of how your business runs. Own that, and which model answers becomes a setting you change on a Tuesday. Own the machine too, and so does the question of whether your work ever leaves the building.
## FAQs
### What actually runs on my infrastructure?
The dashboard, the API server, your Postgres database, the object storage, and the sandboxes that execute agent code. Sandboxes run as local Docker containers rather than in our cluster, and storage speaks the S3 API, so MinIO or any S3-compatible store works.
### What still talks to General Input?
The integration catalog. Operation documentation for tens of thousands of API endpoints is reference data that changes constantly and is identical for every customer, so your deployment reads it from us rather than shipping a stale copy. It describes public APIs, not your data.
### Can the model run locally too?
Yes. Organization custom models point at any OpenAI-compatible endpoint, which is what LM Studio, MLX, llama.cpp, and vLLM serve. When the server is on your network as well, that is just another host on the LAN. No tunnel, no VPN, no public endpoint.
### Does the model ever see my API keys?
No. Credentials are encrypted at rest, injected at runtime inside the sandbox, and stripped before anything returns to the model. That holds whether the model is hosted by a provider or sitting on the desk next to the server.
### What size machine do I need for the model?
It depends on the model, and capacity is the binding constraint. A 27B model at 4-bit is about 17GB and runs on a 24GB Mac mini. DeepSeek V4-Flash is 167GB at full precision and needs a 256GB machine to load uncompressed.