What a Mac Can Run Locally in 2026, and Where the Cloud Picks Up

Apple Silicon turned out to be unusually good at local AI, for one structural reason: the CPU and GPU share a single pool of fast memory, so the whole machine behaves a bit like a graphics card with a generous amount of VRAM. That is exactly the shape language models want.

It also means one number decides most of what you can do. Not the chip generation, not the core count. The memory.

The tiers, and what they actually hold

Our Mac guide lays these out in full, and the short version is that each step up moves you into a different class of model rather than making the same model faster.

8 GB. Small models. Quick questions, summaries, rewrites. Genuinely useful for the everyday, and genuinely limited the moment you ask for something long or structured.

16 GB. The tier where local chat starts feeling good rather than tolerable. This is most people's Mac, and it is enough for the models that handle everyday writing and coding help.

32 GB. Mid-size models, or a chat model and an image model side by side. This is also the floor where small-scale local video becomes a reasonable experience rather than an experiment.

64 GB and up. Enthusiast territory. Large models at heavier compression become realistic.

Intel Macs can technically run small models on the CPU. It is slow enough that it is not worth building a habit around.

For the arithmetic behind these tiers, how much RAM you need for local AI goes through where the memory goes, including the part people forget: the context window takes its share on top of the weights, and that share grows with every token.

A worked example, because the tiers hide the cliff

Take Qwen 3.6 35B A3B. It is a mixture of experts model with about 3B parameters active per token, which reads like a small model and is not one. At the 4-bit quantization almost everyone uses, it is roughly 18 to 20 GB of weights before context.

Run that against the tiers. On 16 GB it does not fit, no matter how few parameters are active. On 32 GB it fits and runs at a slower token rate than you would get from a 24 GB graphics card. There is no gentle degradation between those two outcomes, which is why so many people install a runner, hit a load failure, and assume they configured something wrong.

Now take Hermes 3 405B. At 4-bit, 405 billion weights are hundreds of gigabytes in memory. No Mac configuration reaches that, and none will. That model is a data center job by construction.

The pattern: the memory tier does not make big models slow, it makes them impossible. That cliff is the thing worth planning around.

The two walls a laptop does not climb

Model size. Covered above, and it is the obvious one.

Throughput on media. Images are fine on modest hardware. Video is where a laptop stops being a good idea, and it is not about capability so much as time. A clip that takes a few seconds on a rack of GPUs takes long enough on a laptop that you stop iterating, and iterating is the entire creative process. Our roundup of local AI video generation in 2026 is honest about where that line sits.

There is a third wall that is not about hardware at all: there is no native Mac build of the LU Labs desktop app. It ships for Windows and Linux. On a Mac, the local route means LM Studio, Ollama or a similar runner, and the LU Labs route means the browser.

What a Mac is genuinely better at

This is not a piece arguing you should stop running things locally. Three things a local Mac does that no cloud can:

It works with no connection. On a plane, on hotel Wi-Fi that hates you, in a building with bad reception.

Nothing leaves the disk. Not as a policy, as a physical fact. For work code, health questions, or a journal, that is a different category of assurance from any promise a company can make.

It costs nothing per use. Once the model is downloaded, a long session and a short one cost the same, which is zero.

Where the cloud picks up

The gap is specific rather than general: the models that do not fit, and the media work that is too slow to iterate on.

LU Labs Cloud covers both from a browser tab. The Hosted plan is €19 a month with 900,000 credits and carries 15 chat models, including ones far larger than a Mac can hold, plus 10 image models and 5 video models with clips of 5 or 8 s. Masked editing, background removal, an object eraser and upscaling at 2k, 4k and 8k sit in the image studio, and Hosted includes 2 LoRA trainings a month.

If a month is more than you want to commit, a credit pack needs no subscription: €5 for 165,000 credits, €10 for 350,000, €25 for 900,000, any of them from your first purchase, and the credits never expire. One credit is $0.00001, Flux Schnell is 300 credits an image and Flux 2 Dev is 1,200, so the €5 pack is about 550 quick images or about 137 on Flux 2 Dev. The plans and packs are here.

The work runs on managed NVIDIA H100, A100 and B200 class GPUs, which is why an Intel MacBook and an M-series one perform the same on this route. Your Mac is the screen.

The split most people end up with

Nobody who has tried both runs only one. The settled arrangement looks like this: a local model on the Mac as the default for everyday chat, because it is free, private and offline, and the cloud for the two jobs the laptop cannot take, which are the large models and the media studios.

The setup for the second half is in how to use LU Labs Cloud on a Mac, and the comparison with a local runner is on the LM Studio alternative for Mac page. Cloud is shared hardware, so a request can queue for a few seconds at peak, credits are one pool for text, images and video, and the catalog is curated rather than open ended. Account data is hosted in the EU and you can delete it yourself. We never train on your data and we never sell it. Content involving minors is refused everywhere, always.

Before you buy a Mac for local AI, check the memory number against the model you actually want, not against the model you read about. Start on Hosted if the number does not add up.