What a Mac Can Actually Run Locally in 2026
Every Mac buying guide for local AI leads with the chip. M3, M4, Pro, Max, Ultra. That is the wrong number to look at. On Apple Silicon there is no separate VRAM: the GPU and the CPU share one pool of unified memory with macOS, your browser, your editor, and the twenty tabs you forgot about. A model has to fit in whatever is left over. So the spec that decides what you can load is the memory tier of the machine, and everything else decides how fast the tokens come out.
Practical version: an M4 with 16 GB loads exactly the same models as an M2 with 16 GB. It answers faster. It does not answer bigger.
The Tiers, By RAM
Rough working budget: assume macOS plus a browser eats 4 to 6 GB before you start, and that a model needs its weight file plus 1 to 2 GB of overhead for context. Everything below is 4-bit unless stated.
| Unified memory | What fits | Verdict |
|---|---|---|
| 8 GB | Small models only, in the 4B to 7B range at aggressive quants | Not a serious local AI machine. Use the browser |
| 16 GB | 7B to 12B comfortably, a 14B if you close everything else | The floor. Fine for a private assistant, drafts, summaries |
| 24 GB | Adds the 20B class and a 27B dense model at a low quant | An awkward middle. Better than 16, short of the good stuff |
| 32 GB | Qwen 3.8 27B at Q4_K_M, which is 17.1 GB of weights resident | The first tier we would actually recommend |
| 48 to 64 GB | The 70B class at 4-bit, with room for long context | Serious. Mac Studio and high spec MacBook Pro territory |
| 128 GB and up | Very large mixture of experts models at low quants | Possible, slow, and a lot of money for a niche |
Note what is missing from that table: a 405B model. There is no configuration of laptop that runs one at usable quality, and no quant that changes the arithmetic.
The 32 GB Threshold, And Why We Are Blunt About It
If you are buying a Mac and local AI is part of the reason, 32 GB is the number. Below it you are permanently one model generation behind: the interesting open weight releases keep landing in the 27B to 35B range, and a 16 GB machine cannot hold them no matter how patient you are. At 32 GB, Qwen 3.8 27B fits at Q4_K_M with room for context, and a 0.9 GB companion file gives it image input. Our Qwen 3.8 27B setup guide has the quant table if you want the exact file sizes.
Apple charges a lot for that step. Pay it once, or plan on the hosted lane covering the gap. Those are the two honest options, and buying 16 GB while intending to run 35B models is not one of them.
If you are starting from zero on a Mac, LU Labs has a walkthrough on how to run AI models locally on your Mac, from picking a model size to the first chat.
The Half Nobody Mentions: Diffusion
Text generation on a Mac is genuinely good. Memory bandwidth on Apple Silicon is high and the MLX tooling has matured. Image and video generation is a different story, and this is where the marketing quietly stops.
Diffusion is compute bound rather than bandwidth bound, and it is the workload NVIDIA hardware was tuned for over a decade. The Mac tooling also lags: most new samplers, ControlNets and video pipelines land on CUDA first and arrive on Apple Silicon months later, if at all. In practice a single high resolution image that finishes in seconds on rented hardware can take minutes on a laptop with the fans at full tilt, and a video clip is worse by an order of magnitude. If you want the local diffusion route anyway, our Flux local guide covers what it costs in hardware.
Update for 3.0.6: the desktop app now makes images and video on the Mac itself, and we have timings from one M5 Pro with 48 GB. They are in the setup section below. Version 3.0.7 adds character training, Talking Character, Motion Control and Music on the Mac, listed in the same section. The general point about smaller Macs still holds.
This is where the cloud earns its keep, and we would rather say that plainly than sell a Mac workflow that does not exist. The hosted lanes at LU Labs run on managed NVIDIA H100, A100 and B200 class GPUs, with image models, editing and upscale, and video models producing 5 or 8 second clips. The hardware is shared, so at busy hours there is a short queue.
Where the Line Falls in Practice
| Task | Local Mac | Hosted |
|---|---|---|
| Short chat, drafting, summaries | Yes, and it should stay local | Only if you want a bigger model |
| Long agent loops with tool calls | Painful, and it heats the machine for an hour | Yes |
| A trillion parameter model | No | Yes, Kimi K3 |
| One image, for fun | Workable, expect minutes | Yes, and it is seconds |
| A batch of forty variations | No, this is an evening of waiting | Yes |
| A 5 second video clip | Not really | Yes, LTX-2 and four other models |
| LoRA training | Since 3.0.7, a Z Image character, slowly: the Standard setting runs a little over two hours on an M5 Pro | 2 a month on the Hosted plan, none on a credit pack |
What the Other Side Costs
Concrete anchors, because "cloud is expensive" is usually said without numbers. Flux Schnell is 300 credits per image. Flux 2 Dev is 1,200 credits per image. A 5 second LTX-2 clip is 8,000 credits. A 15 EUR pack is 450,000 credits, needs no subscription, and the credits never expire, which works out to roughly 1,500 quick images or 375 Flux 2 Dev images. The Hosted plan is 19 EUR a month with 900,000 credits, about 3,000 quick images or 750 Flux 2 Dev images, plus the LoRA training slots. Per model rates are at https://lu-labs.ai/pricing.
Compare that against the Apple price of stepping from 16 GB to 32 GB and it is not a close call for image work.
Setting Up Locally Uncensored on a Mac, Step by Step
Our free open source desktop app has a Mac build since version 3.0.6. It is in beta, it needs Apple Silicon and macOS 14 or newer, and Intel Macs are not supported. The earlier sections of this article were written before that build existed. This one covers the setup.
- Download the .dmg for Apple Silicon.
- Open the .dmg and drag the app into Applications.
- Start it. The setup wizard picks a first model for your machine.
- Send a first message in Chat. Local models run through the bundled server with Metal, so there is no separate runner to install.
- Open Create when you want a picture. On a Mac, Create makes images and video on the Mac itself, without ComfyUI.
If you keep models in a folder under Desktop, Documents or Downloads, LU says up front that macOS will ask once for access to it.
What to Expect From Create on a Mac
We have numbers from one machine, an M5 Pro with 48 GB. DreamShaper XL Turbo made an image in 33 seconds and Qwen-Image 2.1 Turbo in 38. LTX-2.3 Distilled 4-bit made a clip of two seconds at 768x512 with sound in 64 seconds. A smaller Mac will be slower, and we have not measured by how much.
Three things the app does for you there:
- A Mac has one pool of memory for chat and Create. Before a picture or video run, Create takes the chat model out of memory, shows "Freeing memory..." while it does, and brings the model back when the run is over
- Where we timed a model on a real run, Create says before the start how long a clip takes once that is ten minutes or more. During the run the line under the bar names the step and the time left
- The 8-bit and Turbo versions of Qwen-Image 2.1 ask for 24 GB. Their runs on our 48 GB Mac peaked at 10.3 and 14.9 GB. We have not run either on a Mac with 24 GB
What 3.0.7 Adds on a Mac
Version 3.0.7 brings more of Create to the Mac, still without ComfyUI. Everything in this list is for macOS only:
- Character Studio trains a Z Image character from 4 to 100 photos. The first run downloads a training adapter of 340 MB. One step takes about 7 seconds on an M5 Pro, so the Standard setting (1200 steps) runs a little over two hours. The trained character then works in Image, the video tabs and the chat tools
- Talking Character turns a portrait and a voice recording into a clip of up to 10 seconds with LTX-2.3 4-bit. It needs one more download of 11.3 GB and a Mac with 24 GB or more, and it is slow: a 4 second clip took about 10 minutes on an M5 Pro
- Motion Control makes a start image follow a motion clip, with LTX-2.3 4-bit and a control file of about 650 MB. A 2 second clip at 512 x 512 took 103 seconds on an M5 Pro
- Music runs ACE-Step 1.5, a download of about 10 GB, for tracks from 10 seconds to 4 minutes
- Enhance Image runs with SeedVR2 3B (7.28 GB) at 2x or 4x. Remove Background uses the subject detection built into macOS and needs no download
- Edit with Qwen-Image 2.1 can paint under a mask and takes up to ten images, as many as the memory holds at the size you set
- HiRes fix runs as a second pass with a model that edits, and Extend Video is a local tab while a video model that takes a start image is installed
All of this ran with the real engine on one Mac with 48 GB. None of it has been clicked through in a built app yet, nothing has run on a Mac with less memory, and training has only run with 10 photos and 100 steps.
Most video models in the Mac catalog are marked "Experimental: not yet tested in LU". Take that label at its word.
What Is Not Confirmed on a Mac Yet
The app has run on that one Mac only. It has not run on macOS 14 or 15, and not on a Mac with 16 or 24 GB. Voice is in the Mac build and has been tried by hand in the app, in English only. Computer control is in the app but not fully tested yet on macOS. Remote Access and the automatic update are not confirmed on macOS. The Mac page keeps this list current, and the 3.0.6 release notes have the rest.
If your Mac is an Intel model or short on memory, the browser Studio covers what the machine cannot hold. The LM Studio alternative for Mac guide walks that setup step by step, including the API key path so the hosted models answer from your terminal too.
If you want to compare the local runners themselves first, the best local AI apps of 2026 roundup covers which ones have decent Apple Silicon support.
Locally Uncensored is AGPL-3.0 licensed and built by PurpleDoubleD. Corrections welcome on GitHub Discussions or in the Discord.
In the handbook
- Locally Uncensored for Mac, the Apple Silicon build, in beta
- Install and update, the SmartScreen warning and how updates arrive
- Chat, the model picker, the context window and thinking mode
- FAQ and glossary, answers to the questions people really ask