Can your computer run local AI? The calculator, with the math shown
Pick your Mac or GPU, get the exact models it runs, with the math shown. Learn the subtraction once and never guess again.
Every local AI forum is one question wearing a hundred outfits. “Can my MacBook Air run DeepSeek?” “Will an RTX 4060 handle a 14B model?” “Is 16 GB enough?” The answers exist, but they are scattered across spec sheets, model cards, and year-old Reddit threads, so the question gets asked again every hour, and answered with a shrug: “depends on your setup.”
It doesn’t depend. It computes. Whether a model runs on your machine is arithmetic: three numbers added up, one comparison at the end. No benchmark required, no downloading a 20 GB file to find out the hard way.
This page does the arithmetic for you. Pick your machine in the calculator below and get the honest menu: which models run well, which barely squeeze in, and which are out of reach. And because every verdict shows its math, you leave with the rule, not just the answer.
The answer is a subtraction, not a spec sheet
One correction before the tool. The instinct is to ask whether your machine is fast enough, and it points at the wrong spec. Speed decides how quickly words appear. Memory decides whether the model loads at all. A model that doesn’t fit in memory doesn’t run slowly; it doesn’t run. We walked through the full logic in the RAM and VRAM explainer; this page turns that logic into a tool.
Knowing that, the shape of the answer is fixed: work out how much memory the model needs, work out how much your machine can actually give it, and compare. The first number comes off the model’s download page. The second comes off your spec sheet, with one platform-specific catch we’ll get to.
And the stakes are ordinary, not exotic. The most common amount of system RAM on Steam’s June 2026 hardware survey is 16 GB, on 41.6% of machines, and the most common graphics card is a laptop RTX 4060 with 8 GB. That ordinary machine runs real local AI today. Not the biggest models, but a genuinely useful menu, and the calculator will show you exactly which rows are yours.
Enter your machine, read your menu
Three inputs: what kind of machine, how much memory, and how much text you plan to feed it in one go. The context input matters more than people expect; a model that fits comfortably for short chats can blow past your memory when you paste a long contract into it, and the calculator makes that visible.
A reading tip: the best pick from your menu is rarely the biggest model that fits. The biggest one saturates your memory, leaves no room for context, and runs the slowest. The daily-driver move is to choose the smallest model that does your job well, then spend the leftover memory on longer context and speed. A 9B all-rounder that answers in two seconds beats a 27B heavyweight that makes you wait, most hours of most days.
Flip through the presets and one pattern jumps out: the menu grows in cliffs, not a slope. Going from 16 to 24 GB on a Mac unlocks the 14B tier. Going from 12 to 24 GB of VRAM unlocks the 27B and 32B tier. And the 70B class stays locked until roughly 48 GB of usable memory, which is why it effectively belongs to high-memory Macs and multi-GPU rigs. Where those cliffs fall is the single most useful thing to know before spending money on hardware.
Every verdict is three numbers and a compare
Nothing in the calculator is a black box. Each row prints its own arithmetic: weights + context + overhead, measured against your usable memory.
Weights are the model file itself, and the sizes are real download sizes, not estimates from parameter counts. At the 4-bit precision most people run (quantization, the trick that shrinks a model to a quarter of its trained size), Llama 3.1 8B is a 4.9 GB file and the 70B is 43 GB. Both land near 0.6 GB per billion parameters, which is the rule of thumb hiding inside every row.
Context is the tax for the text in play. To keep a conversation going, the model holds a working memory called the KV cache, and it grows in a straight line with the length of the text. At 4K tokens of everyday chat it rounds to almost nothing. At a 32K document it costs gigabytes. At a full 128K window it can outweigh the model itself, which is the quiet reason a setup that worked for weeks falls over the day you feed it a book.
Overhead is a flat half-gigabyte for the runtime and its buffers. Small, but real, and honest math includes it.
The verdicts come from the comparison. Under 85% of your usable memory earns runs great: room for longer chats and background apps. Between 85% and 100% is a tight fit: it loads, with no slack. Over 100% is won’t fit, and no setting rescues it. The speed estimates next to each verdict follow from one physical fact: to produce a token, the machine reads the model’s active weights once, so tokens per second is roughly memory bandwidth divided by those bytes. That’s also why the mixture-of-experts coding model in the list posts startling speeds: it stores 19 GB but reads only about 2 GB per token.
That mixture-of-experts wrinkle cuts the other way for memory, and it trips people up on release days. An MoE model activates only a few billion parameters per token, but it stores all of them, so for the fit question you count the full download, not the active slice. Cheap to run, expensive to hold. When a launch post brags that a new model “only uses 3B active parameters,” check the file size before assuming it belongs on a laptop.
A Mac counts to 75%. A PC counts the sticker.
The calculator asks Mac or PC first because the two platforms count memory differently, and this is where most back-of-envelope answers go wrong.
On an Apple Silicon Mac, the processor and the graphics cores share one pool of unified memory, so your RAM is, quite literally, your video memory. The catch is that macOS holds a reservation for itself: by default the GPU’s working set is capped, per Apple’s Metal working-set limit, at roughly three quarters of RAM. That is the 75% haircut the calculator applies: a 16 GB Mac behaves like a 12 GB card, a 64 GB Mac like a 48 GB one. Power users can raise the cap with a one-line override, but 75% is the honest planning number.
A PC with a discrete GPU is stricter, in both directions. Only the VRAM soldered onto the card counts; the 64 GB of system RAM in the same tower contributes nothing at full speed, because any weight that spills out of VRAM has to cross a slow bus on every single token. A model that half-fits doesn’t half-run, it grinds. So the calculator takes the sticker number at face value: NVIDIA’s current lineup runs from 8 GB on an RTX 5060 to 32 GB on an RTX 5090, with the previous generation’s 24 GB cards still very much in the game. In exchange, discrete cards bring far more memory bandwidth, which is why the same model that fits both machines generates several times faster on the GPU.
One trap on the PC side: laptop GPUs wear desktop names with smaller memory. A mobile card often carries less VRAM than the desktop card of the same number, and gaming laptops are marketed on frame rates, not memory, so the spec is easy to miss. If you’re running the calculator for a laptop, look up the VRAM of your exact variant and use the Custom field rather than trusting the model name on the lid.
A tight fit has two escape hatches
A tight fit verdict is not a no. It means the total lands within your budget but leaves little headroom, and you have two dials to turn before giving up on a model you want.
The first dial is precision. The calculator prices every model at 4-bit, the sensible default, but most models also ship smaller builds: a 3-bit or aggressive 4-bit variant typically shaves 15 to 25% off the file for a modest quality cost. That is often exactly the margin between tight and comfortable.
The second dial is context. Switch the calculator between 4K and 32K and watch the verdicts flip: a 14B model on a 16 GB Mac is a tight but workable daily driver at chat lengths, and a non-starter with a 32K document loaded. If you need long documents on small hardware, run a smaller model, or halve the context tax with an 8-bit KV cache, which most runtimes now expose as a checkbox.
The same logic explains the most common local AI mystery: a model that loaded fine for weeks suddenly runs out of memory. Nothing broke. The conversation got longer, the context tax crossed the line, and a tight fit became no fit. Now you can see the line it crossed.
If the menu is too short, buy memory, not speed
If you flipped through the presets and your menu feels thin, the upgrade math falls straight out of the same formula. Memory buys new capabilities; speed only polishes the ones you have. A faster chip with the same memory runs the same list of models, slightly quicker. A slower chip with more memory runs models the fast one cannot load at all.
On the Mac side that decision happens once, at checkout, because unified memory is soldered. The jump that matters most is off the base tier: 16 to 24 or 32 GB turns a chat-only machine into one that runs 14B-class models with room to breathe. Our Mac buying guide walks the whole lineup, but the one-line version is: buy the memory, not the machine.
On the PC side the lever is the card, and the sticker to chase is VRAM. The value play in 2026 is still a used 24 GB RTX 3090 or 4090: an older card with more memory beats a newer card with less for this workload, because 24 GB clears the cliff where 27B and 32B models live. Whatever you do, don’t pay a premium for a 12 GB card on the strength of its benchmark scores; the models it can’t hold don’t care how fast it would have run them.
Never ask “can I run it” again
The calculator is the answer for today’s models, and the rule it keeps printing is the answer forever: multiply a model’s billions of parameters by 0.6 for the 4-bit file, add a gigabyte or two for context and overhead, and compare against your usable memory, which is 75% of RAM on a Mac and the VRAM sticker on a PC. If your number is bigger, it runs.
Models will keep changing; the subtraction won’t. Bookmark this page for the next release day, share your result with the URL (your selection is saved in it), and when a new model drops, don’t ask the forum whether your machine can run it. Read the download size and answer in your head, before the download finishes.
