Open models, closed models, and what your computer can run.

Strategy, not charity. Meta, Google, and the rest release open models so the world builds on their technology instead of a rival's, and every improvement the community makes flows back around.

There's also talent and trust: researchers prefer companies that publish, and businesses trust tools they can inspect and keep. None of this makes the models less real or less useful to you. It just explains why billion-dollar software sits on a free shelf.

Almost. Truly open source would mean publishing everything: training data, code, the recipe. Most open models publish the finished brain (the 'weights') plus a license letting you use it, and keep the recipe private. Purists say 'open weights'.

For running AI at home, the difference changes nothing: you download the file, it works, it's free. A business building ON a model should skim the license once; using one at home, you'll never notice.

Neither. Closed models are smarter; open models are yours. The top closed models win on the hardest work, and the freshest features land there first. (A few giant open models get close, but they need server hardware.)

Open wins on privacy (nothing leaves your machine), cost (free forever once you own the computer), and control (nobody can retire your favorite model). Many people use both: cloud for hard problems, local for private notes and drafts.

Your conversation is sent to the company's servers, processed, and stored under its privacy policy. The big providers let you opt out of training, and business plans usually promise no training at all. Real protections, but they're policies: promises, not physics.

A local model is physics. There's no server, so there's nothing to trust. For everyday questions the difference rarely matters. For a therapist's notes, a lawyer's drafts, or unreleased numbers, it can matter a lot.

Easier than setting up a printer. Download LM Studio (free, Mac and Windows), open it, click download on a model it recommends for your machine (it checks your memory and warns you off models that won't fit), and chat. Most of the time is the download.

No command line, no code, no account. And if you'd rather see it before you try it, that's an easy thing to do together in a session.

LM Studio if you want an app with buttons: a chat window, a model store, and clear warnings when a model is too big for your machine. The right first stop for almost everyone.

Ollama if you want other apps to use your local model as a building block; it's the plumbing many local AI tools connect to, with a simple app of its own these days. You can run both; the model files work the same.

Standard compressed ('quantized') versions, rounded to stay honest. Longer conversations need extra working space, so 'roughly' is doing real work here.

Parameters are the tiny numeric dials inside the model, tuned during training; together they store everything it knows. More dials generally means more capable.

Nobody needs to picture 8 billion of anything. Read 'B' as a size tag: 4B small, 8B everyday, 30B serious, 70B heavyweight. The tag maps straight to memory needed, which is why this whole page runs on it.

Compression for AI brains: storing the dials with less precision, the way a paperback shrinks a hardcover without losing the plot. The standard 'Q4' versions keep nearly all the smarts at about a quarter the size.

It's the form virtually everyone runs at home, and what this page's numbers assume. Apps like LM Studio pick a sensible version automatically; if you're ever offered a choice, take the one marked 'recommended'.

To produce each word, the computer sweeps through essentially the entire brain, billions of numbers, again and again. If the whole brain sits in fast memory, the sweep is quick. If it doesn't fit, speed falls off a cliff.

The processor sets how fast the sweep goes; memory decides whether it's possible at all. That's why this page asks about gigabytes, not gigahertz. (The 'NPU' chips in ads mostly accelerate small background features; see the questions below.)

On a typical PC, big models must fit inside the graphics card's own small memory pool: 8, 12, maybe 24 GB on expensive cards. Apple-chip Macs share one big pool of fast 'unified memory', and the model can use about two-thirds to three-quarters of it.

So a 64 GB Mac can hand roughly 45 GB to a model, something a PC needs exotic hardware to match. That's why Macs became the darling of the local AI crowd. (The PC world is catching up: a few 2026 machines borrow the same shared-memory idea, some with up to 128 GB.)

AI speed is measured in tokens per second; a token is roughly three quarters of a word. People read at about 4 to 5 words per second, so past roughly 6 tokens per second a model feels fluid: it's outrunning your eyes.

An everyday 8B model on a recent Mac or a decent graphics card clears that bar several times over. The slow experiences are mismatches: a model too big for the machine, or a processor doing a graphics card's job. The checker below keeps you out of those.

Run the checker above on the computer you already own. If it lands in any tier that runs models, try local AI free for a week before buying anything.

Most people discover the machine they have is enough to know whether they even like this.

A Mac mini with 16 GB of memory: the least expensive machine we'd honestly call good at local AI. The everyday 8B class plus OpenAI's gpt-oss-20b, quietly, on a desk.

Apple's unified memory does the work of a graphics card here. One warning: the cheapest Mac in the lineup (the $699 MacBook Neo) has 8 GB and is NOT this machine.

A Mac (mini, Air, or Pro) with 24 to 32 GB of memory, or a desktop PC with a 16 GB NVIDIA card. This opens the 20 to 32B class, where local AI gets genuinely impressive.

The 32 GB Mac route is simplest and quietest; the PC route also buys you gaming and upgrade paths.

A Mac with 64 GB or more (Mac Studio or MacBook Pro), a PC with a 24 GB card, or one of the new 128 GB shared-memory PCs. This is 70B territory, and at the top end, gpt-oss-120b territory.

Only worth it when privacy-critical serious work is the plan. Past this point you're building a small home lab, and we'd love to hear about it.

Yes, genuinely. Recent iPhones and Android flagships run the 1 to 4B class through free apps: useful for summaries, quick drafts, and offline questions.

Temper expectations: phone models are the pocket-knife version, and long sessions warm the phone. A neat, private convenience, not a ChatGPT replacement.

Only once, for the download. Model files are big (a few GB small, tens of GB large), so decent Wi-Fi helps that first fetch.

After that, everything runs with the network unplugged. No connection, no account, no pinging home. That's the point.

Fully legal. The companies publish them on purpose, with licenses that allow personal and business use; a few add conditions that only kick in at giant-company scale.

The honest catches are practical: models are only as current as their training date, they make mistakes with confidence, and the downloads are big. If a business use rides on one model, skim its license once; it's a page, not a book.

No. Every model only knows what it learned in training. Cloud tools get around this by quietly searching the web before answering; a plain local model can't.

Some local apps bolt on web search, but it's tinkery. If your daily use is 'what's happening right now', cloud fits better. Local shines for writing, thinking, and your own documents.

Yes. Image generation has its own open models (Stable Diffusion and Flux are the famous names) and free apps, and the same rule applies: more memory, better models.

It's a deeper rabbit hole than local chat: more knobs, more jargon. If your machine scored well in the checker, it can likely make images too.

An NPU is a small chip for light AI features: background blur, live captions, photo search. Nice to have, nearly irrelevant to the chat models this page is about, which lean on the graphics side and, always, on memory.

Treat 'AI PC' stickers as marketing. The questions that matter stay the checker's: how much memory, and how fast?

Sometimes, when privacy is the driver: client files, medical or legal notes, anything that shouldn't leave the building. One capable Mac can be a private AI workstation for a small team, monthly cost zero.

The trade is capability and convenience: cloud tools are smarter, current, and need no babysitting. Many businesses use both: cloud for general work, local for the sensitive pile. Sorting which is which is exactly what a setup session is for.