The best Mac for local AI in 2026: buy the memory, not the machine

On June 25, 2026, Apple did something it almost never does: it raised the price of nearly every Mac, mid-cycle, with no new hardware attached. Tim Cook called the increases “unavoidable” and compared the memory-chip shortage behind them to a “hundred-year flood.” AI data centers are buying up the world’s DRAM, and the bill has reached the laptop aisle.

Quieter than the price hikes, and more interesting, is what disappeared. The Mac Studio’s 128 GB, 256 GB, and 512 GB memory options were pulled from the configurator this spring; the Mac mini lost its 64 GB tier. As of July 2026, the most unified memory you can order in any Mac is 128 GB, and it comes in a MacBook Pro. Apple’s biggest-memory computer is a laptop.

That accident of supply chains is the perfect frame for a buying guide, because if you want to run AI models on your own machine, unified memory is the whole decision. It determines which models you can run at all, it is soldered to the package so you can never add more, and it just became the scarcest thing in the lineup. Buy the memory, not the machine: pick the RAM tier your ambitions need, then choose the cheapest Mac that carries it.

One spec decides everything, and it just got scarce

A quick sanity check before spending anything: why run models locally at all, when a cloud chatbot is a browser tab away? Three reasons keep coming up. Privacy: your documents never leave the machine. Cost: weights you have downloaded generate for free, forever. And control: a local model cannot be retired, rate-limited, or repriced under you. We have made the longer version of this case in our guide to the best open models, and the short version is that open-weight models crossed the good-enough line for most everyday work about a year ago.

What has not crossed any line is hardware. A language model must sit entirely in fast memory while it runs. A model that needs 40 GB does not run slightly worse on a 32 GB machine; it does not run, or it swaps to disk and produces a word every few seconds. Memory is a cliff, not a slope. That is why this guide is organized around one number, and why the number going scarce matters: the machine you buy this year fixes your ceiling for the five or so years you will own it.

On a Mac, RAM is VRAM

On a PC with a discrete graphics card, the model must fit inside the card’s own VRAM, and VRAM is brutally rationed. NVIDIA’s flagship gaming GPU, the GeForce RTX 5090, ships with 32 GB, and it alone costs as much as a MacBook Pro. Apple Silicon works differently: the CPU and GPU share one pool of unified memory on the same package. There is no separate video memory and no copy step. Whatever RAM your Mac has, the GPU can use most of it.

“Most” has a specific value. By default, macOS lets Metal, the graphics layer local AI apps run on, wire down roughly 75% of physical RAM for the GPU. A 64 GB MacBook Pro is, in effect, a 48 GB graphics card that also happens to run macOS. Power users can raise the cap with a one-line sysctl command on machines with memory to spare, keeping 8 to 16 GB back for the system, but the 75% rule is the honest planning number.

Run the comparison once and the Mac’s odd position in the AI hardware market becomes clear. GPU-usable memory: gaming flagship, 32 GB; MacBook Pro, up to 96 GB by default and more with the override. For the specific job of holding a large model in memory, a well-configured laptop outclasses a desktop graphics card that draws six times the power. What the discrete card keeps is raw speed on the compute-heavy parts, which brings us to the second number.

Memory size decides what runs; bandwidth decides how fast

Generating text is a strange workload: for every single token, the chip reads essentially the entire model out of memory. The bottleneck is not arithmetic but the pipe between RAM and GPU. Tokens per second track memory bandwidth almost linearly, which makes bandwidth the second spec to read on the box, and conveniently, it is the spec that separates Apple’s chip tiers.

Apple’s own machine learning team published the cleanest evidence for the bandwidth rule. Comparing M5 to M4 in MLX, its open-source inference framework, generation speed improved 19 to 27 percent, almost exactly the 28 percent bandwidth increase (120 to 153 GB/s). The M5’s new GPU neural accelerators made prompt processing up to 4x faster, so long documents start answering sooner. But once tokens are flowing, the pipe is the pipe.

The bandwidth rule also gives you a free speed estimator. Divide bandwidth by model size and you get the ceiling on tokens per second: a 4-bit 8B model is roughly 4.5 GB, so a 153 GB/s chip tops out somewhere near 30 tokens per second on it, several times faster than most people read. Put a 40 GB 70B model on the same chip and the same division lands under 4. Real numbers come in below the ceiling, since the machine has other work to do, but the ratio is what matters: to keep a big model conversational, the pipe has to grow with it.

In practice: a base M5 at 153 GB/s runs an 8B model at comfortable chat speed. An M5 Pro doubles the pipe; an M5 Max at 614 GB/s quadruples it, which is what makes 70B-class models pleasant rather than merely possible. The rule of thumb worth memorizing: memory size sets which models you can run, bandwidth sets how fast they talk, and Apple sells the two together. Climbing the chip ladder buys both.

The memory ladder: what each tier actually runs

Model sizes cluster into bands, so each memory tier unlocks a distinct class of capability. The fits below assume 4-bit quantization, the standard compression for local use, and leave room for context. The arithmetic behind them is one subtraction, worked through in our memory-math explainer.

Two rows deserve a note. Apple’s MLX team offers a useful calibration for the 24 GB tier: it fits an 8B model at full BF16 precision or a 30B-class mixture-of-experts at 4-bit while keeping the whole workload under 18 GB. And the 64 GB row is the one to circle: it is the cheapest tier where 70B-class models, the point where open models stop feeling small, fit at all.

The lineup, mapped to ambition

MacBook Air (M5, from $1,299). The March 2026 Air starts with 16 GB and configures to 24 or 32 GB, all at 153 GB/s. A 24 or 32 GB Air is a genuinely good starter local AI machine: quiet, fanless, and able to run 14B to 32B models at usable speeds. It is the right answer for “I want private chat, drafting, and summarization on my laptop.”

MacBook Pro (M5, from $1,999). Same 153 GB/s and same 32 GB ceiling as the Air. You are paying for the screen, the fan, and the ports, not for more AI. If local models are your priority at this price, a maxed Air or a step up to the M5 Pro are both better trades.

MacBook Pro (M5 Pro, from $2,499).Up to 64 GB at 307 GB/s. This is the enthusiast pick: the 64 GB configuration is the cheapest Mac that runs 70B-class models, and the bandwidth doubles the Air’s. Most people reading this guide with serious intent should land here.

MacBook Pro (M5 Max, 14″ from $3,599 at launch). Up to 128 GB at 614 GB/s, with prompt processing 4x faster than the M4 generation. The 128 GB configuration is the current ceiling of the whole platform: 120B-class models run comfortably, and several mid-size models can stay resident at once. It is a workstation that closes.

Mac mini (M4 from $799, M4 Pro from $1,599). The budget desktop route. The M4 Pro mini at 48 GB and 273 GB/s makes a tidy always-on model server for a household or small office, though the 64 GB option is a casualty of the shortage. Still on M4 chips, with M5 minis expected later this year.

Mac Studio (M4 Max from $2,499, M3 Ultra from $5,299). Normally the local AI king, currently the awkward pick. The remaining 36 to 96 GB configurations are capable, and the M3 Ultra’s 819 GB/s is still the fastest memory Apple ships. But the huge-memory tiers are gone, the M3 Ultra’s price rose $1,300 in June, and an M5-generation Studio is expected around October. Unless you need a desktop today, wait this one out.

Buy one tier above your ambition, skip the rest

The happy corollary of memory-first buying is that you can be cheap everywhere else. Storage is the easiest saving: models are big downloads, but an external Thunderbolt SSD holds a model library happily, since weights load into RAM once and stay there. CPU cores barely matter for inference. Even the GPU-core upgrades within a chip tier matter less than stepping up a memory tier. One box on the configurator does the work of all the others.

Buy one tier above what you plan to run today. Models grow, context windows grow, and macOS itself needs its share. The buyer who sized 16 GB for an 8B model in 2024 met the 40 GB models of 2026 with no recourse: on Apple Silicon, RAM cannot be upgraded later, full stop. The upgrade you skip at checkout is skipped for the life of the machine.

Honesty also cuts the other way: do not overspend. If your local AI life is private chat and document work with 8B to 14B models, a 24 GB Air covers it, and the money saved buys years of API credits for the rare heavy job. The 128 GB Max is a professional tool, not a status default. Match the tier to a concrete ambition, and if you cannot name the model you would run on 128 GB, you do not need 128 GB. Once the hardware is home, the Mac apps that put it to work are the easy part.

The shortlist

Private chat and writing help: MacBook Air, 24 GB. Serious daily local AI, 32B models with room: MacBook Air or MacBook Pro at 32 GB. The enthusiast sweet spot, 70B models on a laptop: MacBook Pro M5 Pro, 64 GB. The full workstation, 120B-class and multi-model workflows: MacBook Pro M5 Max, 128 GB. A desktop on a budget: Mac mini M4 Pro, 48 GB. A monster desktop: wait for the M5 Mac Studio this fall, and hope the memory tiers come back with it.

And if the machine is a year or three away, remember the meta-lesson of this strange spring: the AI boom repriced memory across the industry, and Apple’s configurator is now where that scarcity shows up first. The spec that matters is the one under pressure. When in doubt, buy the memory.