Under a video called "Can a Mac Mini run local AI?", a commenter with 5 likes asks the buying question the whole thread orbits: "The question is which is truly needed for local ai. Or is 16 gigs ram enough or is 24 to 32 the preferred." (@mkII., on the Alex Finn walkthrough).
The question outlived that video. Under an "I ran Qwen 3.8 on a laptop" upload from two days before this post, commenters compare what their Mac Minis and gaming PCs handle; a DeepSeek desktop-cluster video the week before collected memory specs in its comments. One question sits under all of it: can the computer you already own do this job, and what does it cost to make it do more?
What local AI is
Local AI is a language model that runs on your own hardware. The model's parameters, downloaded as files, sit on your disk; when you send a prompt, your computer's memory and processor generate the response. With ChatGPT or Claude, your text travels to the vendor's servers, rented GPUs run the model, and the answer comes back. With local AI, that round trip disappears: the same kind of model runs on your machine instead.
The models are open-weight ones, published by labs like Alibaba (Qwen), DeepSeek, and Meta (Llama), free to download, though licenses differ by model. The apps that run them are free too. Two free apps cover most of it: LM Studio, a point-and-click desktop app, and Ollama, which describes itself as "the easiest way to run and build with open AI models locally or in the cloud on macOS, Windows, and Linux."
Why people run models on their own machine
Three reasons cover it, and the threads supply an example of each.
Privacy, in a basic setup. A local model has no server to send your text to. That beats any cloud setting: there's no training toggle to find and no retention policy to trust. The app around the model can still have online features (telemetry, connected tools, web search), so check its settings before handing over anything sensitive. For the commenter who wrote "that's insane to me in Europe that you share all your sensitive data with those companies" (on a Jeff Su video, September 2026), local is the setup that ends the question.
A different cost shape. No subscription and no per-token cloud charge for generation. You pay for hardware once, plus the electricity to run it; after that, queries cost nothing extra.
Availability. Works offline in a basic setup, at any hour, with no vendor rate limit on a single-user machine. For repetitive daily jobs, that reliability is the feature.
The hardware question, answered
Available memory is the first limit, not the only one. A rough starting estimate from ModelFit's local model catalog: at the common Q4 compression setting, a model needs about 0.6 GB of memory per billion parameters, plus memory for your operating system, the app, and the conversation. Their published tiers, which they mark as estimates from quantization math rather than measured benchmarks:
- 8 GB runs small models, up to roughly 7 to 9 billion parameters
- 16 GB runs 14B-class models (a 14B model takes about 8.4 GB at that setting)
- 24 to 32 GB runs 27B- to 32B-class models (Qwen3.8's open 27B needs 24 GB)
- 64 GB reaches 70B-class models
A model that fits can still run slowly on a modest machine: fitting is about memory, not speed.
If your current computer has 16 GB, capable models run on it today, free, and we'd start there before buying anything. If you are buying, match the memory tier to the model size you want, because on Macs the unified memory can't be upgraded later. The current Mac mini starts at $899 with 16 GB of RAM (TechCrunch, August 25, 2026), a price that climbed twice in 2026 as memory costs rose (Macworld). At the high end, NVIDIA's DGX Spark desktop launched at $3,999 in late 2025 and cost $4,699 by February 2026 (per Vast.ai's hardware writeup). Our short answer to the commenter's question: pay for the smallest tier that runs the model your work needs, and no larger.
How to try it in ten minutes
Both tools are free, and neither needs an account to run local models.
LM Studio is the point-and-click route. It's a desktop app for Mac, Windows, and Linux: open the Discover tab, pick a model from the curated list or search, download it, and chat in the same window.
Ollama is the route that grows into automation. It runs as a desktop app or a terminal command, and it connects local models into the tools you may already use: its docs cover wiring models into Claude Desktop or ChatGPT (Desktop), and a command like ollama launch claude starts a coding agent with a local model behind it. The docs recommend 8 GB of available memory as a starting point; with less, models can still run on system RAM, with slower responses (docs.ollama.com).
Start small either way. A 7B- to 9B-class model can download in minutes on a fast connection, and it gives you a quick sense of the experience before any hardware purchase enters the picture.
What local models can't do
Capability is the gap that decides the trade. A model that fits a laptop is weaker than the frontier cloud models on hard reasoning, and the top comment under the Qwen laptop video says it from experience: "Just as capable as frontier -> no they're not" (4 likes), from someone who bought an AI laptop, found even 128 GB of RAM short of what his jobs needed, and returned it within a day.
Context windows run smaller too, because every token of context occupies memory; the tiers above assume default settings, and longer conversations push a model out of bounds faster. And the most capable models are usually closed: frontier labs keep their weights private, so the local catalog trails the best cloud models. The comment threads are enthusiastic about what local models can do, and honest about this: the hardest problems still go to the frontier models.
The hybrid split
Use local models for the repetitive and private, and cloud models for the hard. That's the division the daily users in these threads describe. One commenter runs Qwen on a 48 GB Mac Mini for daily agent chores (browser automation, email checks) precisely to "take the load off frontier models" and stop spending tokens on routine steps. Another keeps routine work on local tools and planning on a cloud model (on an OpenAI announcements video, September 2026).
The split answers the hardware question the post opened with: match the machine to the routine jobs you want off the subscription, keep the cloud model for work that needs the frontier, and upgrade the hardware only when a specific model you want requires the memory. Local AI doesn't replace the best model you can reach; it moves the routine, private jobs off the cloud bill and onto your own machine.
FAQ
Is local AI free?
The software and the models are free to download, though licenses differ by model. On hardware you already own, that makes it free. Costs enter when you buy hardware: around $899 for a Mac mini that runs 14B-class models, several thousand for 70B-class machines.
Can my laptop run AI locally?
Check its memory against the tiers: 8 GB runs 7B- to 9B-class models, 16 GB runs 14B-class. A machine with a dedicated GPU or Apple silicon runs them faster than an ordinary laptop on system memory alone, and the model that fits may still run slowly on modest hardware.
Will a local model be as good as ChatGPT?
Not on hard reasoning: the frontier cloud models are stronger, and the most capable models usually aren't available to download at all. For routine jobs (summarizing, drafting, simple automation), capable open models do the work, which is why the hybrid split exists.
Is local AI better for privacy?
In a basic setup, yes: your prompts and documents stay on the machine, so there's no training toggle to find and no retention policy to trust. Check the app's own online features (telemetry, connected tools) before handing over anything sensitive; a local model with chatty integrations isn't fully local.
Does local AI work offline?
In a basic setup, after the model downloads: generation happens on your hardware, and no internet connection is needed to keep chatting. Downloads and updates are the parts that need a connection.
Want to go further?
Do AI Agents Train on Your Data? What Agents See, Save, and Learn covers what cloud services do with your text, the problem local AI removes at the root. What Is a Token in AI? The Unit Behind Every Price and Limit explains the unit a local setup never bills. The one-step version of this post: if your computer has 16 GB, install LM Studio, download a small model, and try a non-sensitive prompt first; check the app's settings for online features before you give it anything private.