MacBook Air vs MacBook Pro for local AI: fanless vs actively cooled

Disclosure: I build Outlier, a Mac app that runs language models entirely on-device, so I'm an interested party in people buying Macs that run models well. I'll try to be useful anyway — this page is about the hardware, and the honest answer for most people is the cheaper machine.

Memory capacity is the decision. Bandwidth is the experience.

Everything a local model needs — the weights, the KV cache that grows with your conversation, plus whatever macOS and your other apps hold — has to fit in unified memory at once. That's a wall, not a slowdown. If a model doesn't fit, it doesn't run.

The MacBook Air (M5) ships with 16GB and configures to 24GB or 32GB. End of the line. An M5 Pro ships with 24GB and configures up to 64GB, but only on the 20-core-GPU binning, not the base 16-core part. An M5 Max ships with 36GB and goes to 48GB, 64GB, or 128GB, and 128GB only on the M5 Max with the 40-core GPU. So the step to the 128GB class isn't a $249 checkbox. It's the step from a $1,299 Air to a $4,099-and-up M5 Max.

The sticker number isn't your model budget, either. Metal exposes recommendedMaxWorkingSetSize — how much memory a GPU "can allocate without affecting its runtime performance" — and Apple advises staying under it. It sits below installed RAM, so budget a few gigabytes of overhead.

Bandwidth is the other half: 153GB/s on M5, 307GB/s on M5 Pro, 460 or 614GB/s on M5 Max by GPU core count — roughly a 4x spread. llama.cpp's maintainers put the mechanism plainly: token generation is bandwidth-limited, prompt processing is compute-bound. Words on screen track bandwidth; time-to-first-token on a long document tracks compute.

What actually fits, at real model sizes

Gigabytes are hard to reason about abstractly, so here are the tiers I ship in Outlier, which publish RAM floors. Comparable apps have a similar ladder.

• 16GB Air ($1,299): Nano 4B (2.37GB download, 6GB floor) and Lite 9B (5.04GB, 12GB floor) run comfortably. Quick, a 26B mixture-of-experts model at 15.61GB, has a 16GB floor — it fits, with little room for anything else.
• 24GB Air, or a base M5 Pro: where 27B-class dense models get practical. Core 27B and Code 27B are 15.13GB downloads with a 24GB floor; Vision 35B-a3b is 19.0GB.
• 32GB Air (the ceiling): same models, with headroom for long contexts and a full desktop behind them.
• 64GB+, MacBook Pro only: Outlier's Plus tier is a 397B mixture-of-experts model: a 209GB download, 64GB floor, cleared by streaming experts from SSD at roughly 11GB peak resident memory. No Air will ever run it.

Apple says the M5 Max can "run LLMs with hundreds of billions of parameters entirely on device." True, and more than most people need. Storage matters too: 209GB is no rounding error on a laptop SSD. Base Air storage doubled to 512GB, and Apple credits the new SSD with "2x faster read/write performance," which it says is "significantly accelerating file access and speeding up workflows." That's a file-access claim, not a model-load claim, and Apple publishes no figure for how long a local model takes to load on any Mac.

Thermals: what fanless actually costs

Apple says both new Airs "feature a thin, light, and completely silent fanless design." The Pro has active cooling, and on M5 Pro and M5 Max a High Power Mode that "allows the fans to run at higher speeds" and "may allow the system to deliver higher performance in very intensive workloads." Apple's model list for it includes the 14- and 16-inch MacBook Pro with M5 Pro or M5 Max, and no Air and no base-M5 14-inch Pro, so "buy a Pro" only buys that lever on an M5 Pro or Max.

I have to be careful here. I found no rigorous primary measurement of throttling under GPU-heavy inference on a fanless Air. The figures that circulate ("20 tokens/sec falling to 8," "15-25% after half an hour") appear only on affiliate blogs with no methodology. I'm not citing them.

The closest credible data point is CPU-only and a generation old: Notebookcheck's M4 MacBook Air 15 review found that in a Cinebench 2024 multi loop, "after 30 minutes, its performance of 776 points was only 11% lower," down from an unthrottled 874, with package power peaking at 20.5W and settling near 9W after 20 minutes. Third-party, and not an inference test, but the shape is instructive. The Air throttles, gradually, and lands well above zero.

Practically: a chat turn that takes twenty seconds never approaches the thermal limit. Batch-summarizing two hundred documents does. If your use is bursty, cooling barely matters. If it's continuous, that's the whole reason the Pro exists.

Side by side

The per-gigabyte row divides starting price by standard memory, which flatters the Air. I couldn't confirm Apple's 24GB and 32GB upgrade prices from a primary source, since the store renders them client-side. It also uses the base 14-inch M5 Pro, the 15-core CPU / 16-core GPU part, which can't be configured to 64GB at all; reaching 64GB means starting from the 18/20 model instead.

Where each one genuinely wins — and where neither does

The Air wins on capacity per dollar, and on silence. A $1,299 fanless laptop holding a 15GB model in memory the GPU addresses directly, with no VRAM copy, is not a compromise machine. MLX, Apple's MIT-licensed array framework, is built around exactly that: arrays live in shared memory and run on CPU or GPU without transfers. And silence is a real feature. A long generation on a PC laptop announces itself through the fans; on an Air it doesn't announce itself at all.

The Pro wins on ceiling, prompt processing, and sustained load. 128GB at 614GB/s has no equivalent in any other laptop class, and Apple claims the M5 Pro and Max deliver "up to 4x faster LLM prompt processing than M4 Pro and M4 Max" — the compute-bound half, the half you feel when you paste in a long file.

A desktop with a discrete GPU usually wins on raw throughput, and I won't pretend otherwise. If a model fits inside a dedicated GPU's VRAM, that card will typically generate tokens faster than any Mac laptop, because it has far more bandwidth to work with. Apple's advantage is narrower: capacity per dollar, and power draw. If your workload fits comfortably in 16 or 24GB of VRAM and you have a desk and an outlet, a PC with a discrete card is a defensible choice.

Nobody wins on published numbers. Every Apple AI claim is a relative multiplier — "up to 4x faster performance for AI tasks than MacBook Air with M4," "up to 9.5x faster than MacBook Air with M1" — against unstated workloads, and Apple gives no absolute tokens-per-second figure for any Mac. The battery ratings (18 hours on the Air, up to 24 on the 16-inch M5 Pro) are video-streaming tests. And the llama.cpp community table is crowd-sourced rather than vendor-measured, useful as direction and nothing more.

Who should pick which

Buy the Air, and put the savings into memory. If local AI is a tool in your day rather than your whole day — drafting, code assistance, summarizing, private notes — a 24GB or 32GB Air runs 27B-class models and never makes a sound. Outlier's Nano tier ran ~32 tok/s in a single-prompt observation on a 16GB M4 Air (2026-05-04, indicative), quicker than most people read. Skip the 16GB base config if local AI is a real reason to buy.

Buy an M5 Pro at 64GB if you want the very large mixture-of-experts models on a laptop, or if you routinely run inference for tens of minutes at a stretch and want High Power Mode as a lever. Check the chip line before you check the memory box: Apple only offers 64GB on the M5 Pro with the 20-core GPU, so the base 14-inch M5 Pro is a dead end for this configuration.

Buy an M5 Max at 128GB only if you can name the reason: models above the 64GB class, fine-tuning, several models resident at once. Otherwise it's an expensive way to buy headroom you won't reach.

One caveat that isn't hardware at all: whether inference is local is a property of the app, not the Mac. Apple's own Apple Intelligence is hybrid, routing some requests to Private Cloud Compute, where Apple states "PCC deletes the user's data after fulfilling the request."