Uncensored AI on Mac: A Memory-Based Model Selection Guide

Key Takeaway: Check Your Memory First, Then Pick a Model

Running "uncensored AI" on a Mac is really two problems stacked together: which models don't restrict requests, and which models fit into your Mac's memory. The second question is often more decisive than the first.

According to information disclosed by developers on their blog, the Mac version of Locally Uncensored is still in beta, and has only been tested on a single M5 Pro (48GB memory). The size figures in this article come from the model files themselves; scenarios that haven't actually been run on a Mac are clearly noted.

What "Uncensored" Means in Local Models

Local models have no moderation layer in front of them. Nothing checks your prompt before the model reads it, and nothing checks the response afterward. What's left is the model's own training — and for default models from large labs, one of the training objectives is to refuse a range of requests.

What people call "uncensored" models change that. The two most common labels are:

• Abliterated: removes the direction in the weights that produces refusal behavior
• Heretic: abliteration performed with the Heretic tool — same idea, but automated

Neither is magic. An abliterated model can still occasionally refuse to answer, and removing refusal behavior may also cause quality losses in other areas. If a model repeatedly dodges what you're asking about, try another model on the list rather than second-guessing your prompt first.

The app itself adds nothing extra. The default persona simply states a role and contains no directional content rules, so what you get is the model itself.

Abliteration is a post-processing technique applied to model weights. Researchers found that after extensive safety training, large language models develop certain "refusal directions" in their weight space — a small set of activation patterns strongly correlated with refusal behavior. By orthogonalizing or reducing projections along those directions, you can substantially lower the probability of the model triggering a refusal while preserving other capabilities as much as possible. This process requires no retraining, making it far cheaper than fine-tuning from scratch. The side effect is that some model accuracy may be lost, and the model may even become more "chatty" in other safety domains, so quality varies depending on the base model.

MoE (Mixture of Experts) is the architecture used by models like Qwen 3.6 35B. It splits the feed-forward layers into multiple parallel "expert" subnetworks and, during inference, activates only a few experts via a gating mechanism. The result is a large total parameter count (e.g., 35B) but only a small fraction actually participating in computation per forward pass (e.g., 3B), so inference speed approaches a 3B model while capability approaches a larger one. On Mac unified memory, MoE still needs to load all weights into memory; fewer active parameters only affect computation, not memory footprint — which is why a 35B MoE still requires 48GB-class memory.

The Key Number: Unified Memory

On Apple Silicon, there is no dedicated VRAM. The model shares the same memory pool with macOS, the browser, and every other program. The chip name determines response speed; the memory size determines which model you can load. An M4 with 16GB and an M2 with 16GB can load the same models.

A practical rule of thumb: the model file plus 1–2GB of context overhead must fit into the space remaining after macOS and open apps consume 4–6GB.

Unified Memory is a core architectural feature of Apple Silicon: the CPU, GPU, and neural engine share the same physical memory, with no separate VRAM layer. This leads to two differences from traditional PCs. First, model weights exist in memory as a single copy, and passing tensors between CPU and GPU requires no copying over a PCIe bus, making inference more efficient. Second, the memory available to a model effectively equals total memory minus what macOS itself and all open apps consume — so a 16GB Mac with a browser and IDE open often has only 8–10GB free, whereas a discrete GPU with 16GB of VRAM can use nearly all of its capacity exclusively. This is the frequently underestimated bottleneck in Mac model selection, and the reason this article repeatedly stresses that "memory matters more than the chip model."

Chat Models Matched to Memory

The uncensored chat models recommended in the app's catalog, ordered by download size:

Model	Size	Characteristics
Llama 3.1 8B Abliterated	5 GB	Usually the first download
Qwen3-VL 8B Abliterated	5 GB	Can read images
Gemma 4 12B Heretic	7.4 GB	Can read images
GLM 4.7 Flash Heretic	10 GB	A 30B-class model in a small build
Qwen 3.8 27B Uncensored	18 GB	Images, tools, switchable thinking
Qwen 3.6 35B MoE Abliterated	24 GB	35B total parameters, 3B active

Applying the rule above, the rough correspondence is:

• 16 GB Mac: Can run 5GB models fairly comfortably; Gemma 4 12B Heretic requires fewer background apps
• 24 GB Mac: Adds GLM 4.7 Flash Heretic
• 32 GB Mac: Adds Qwen 3.8 27B Uncensored — this is the recommended minimum tier
• 48 GB and up: Adds Qwen 3.6 35B MoE Abliterated, with room to handle long context

This mapping is purely arithmetic based on file sizes. The author has only run chat functionality on a 48GB machine, so performance on 16GB or 24GB Macs should be treated as a starting point, not a verdict.

The setup wizard will pick the first model for your machine, and the model manager will flag which ones fit. You don't need to do these calculations yourself, but understanding the reasoning helps you make sense of the options the app offers.

Image and Video Generation

The Create feature on Mac can generate images and videos locally, with no ComfyUI required. The Mac catalog has a separate group called Unfiltered, alongside Mainstream:

• Image models: 23 total in the catalog, 10 of them in the Unfiltered group
• Video models: 31 total in the catalog, 14 of them in the Unfiltered group

Each model card states its required memory, and the app enforces it strictly. Image models require a minimum of 8GB; video models require a minimum of 16GB. Most video models in the Mac catalog are marked "Experimental: not yet tested in LU" — that label means exactly what it says.

Speed reference: on an M5 Pro with 48GB, DreamShaper XL Turbo generates an image in 33 seconds, and Qwen-Image 2.1 Turbo takes 38 seconds.

ComfyUI is currently the most popular local AI image/video generation workflow software, known for its node-based interface — powerful but complex to configure, typically requiring users to install PyTorch, download checkpoints, and manage LoRA and ControlNet themselves. Locally Uncensored builds this kind of capability into the app, saving you the step of setting up ComfyUI, at the cost of limiting model selection to the app's catalog and offering less extensibility than ComfyUI. The Turbo suffix (as in DreamShaper XL Turbo, Qwen-Image 2.1 Turbo) usually refers to model versions that use distillation to compress sampling steps, producing images in 4–8 steps rather than the 25–50 steps often needed by non-Turbo versions — so Turbo is several times faster on the same hardware, at the cost of slightly reduced detail and diversity. This is why the generation times mentioned here stay in the 30–40 second range.

Installation and Running

System requirements: Apple Silicon, macOS 14 or newer. Intel Macs are not supported.

1. Download the .dmg file from locallyuncensored.com/mac — it's signed with a Developer ID and notarized by Apple
2. Drag the app into the Applications folder and launch it
3. Let the setup wizard pick the first model, or open Models and choose from the list above
4. Send a message

No account or login is required for local use.

What Data Gets Sent

When using local models, your prompts and responses stay on your machine. The app has no analytics package and sends no crash reports. Two pieces of data are sent out automatically: a query to GitHub Releases for new versions a few seconds after launch and then at most every six hours; and a count each time the Cloud switch is toggled (direction, platform, and app version only, no identifiers). If you never touch the Cloud switch, the second one never fires.

What's Not Yet Available on Mac

• Training your own character LoRA cannot run locally on Mac
• Computer control exists in the app but is not fully tested on macOS
• Voice features work locally but are in beta and English-only
• Remote Access and auto-update are unconfirmed on macOS

Brief Summary

Uncensored AI on Mac comes down to picking an abliterated model that fits your unified memory. 16GB handles the 8B class, 32GB is the fit point for Qwen 3.8 27B Uncensored, and the app's Unfiltered group covers images and video on the same machine. Everything runs locally with no account required — but this version is still in beta and has only been verified on a single Mac.

If your Mac doesn't have enough memory for the model you want, larger models can also be hosted in the browser via LU Cloud, which is a paid service; all the local features above are free.

One-line Summary

To run uncensored AI on a Mac, first check whether your unified memory can hold the model — then worry about whether the model holds back.

Paragraph Summary

The article notes that using "uncensored AI" on a Mac is really the overlap of two problems — memory and model — and memory is often the more decisive one. Apple Silicon uses a unified memory architecture, so models share one memory pool with the system and apps: the chip model determines speed, while memory size determines which models can load. The article offers a practical estimate: a model file plus 1–2GB of context overhead must fit into the space left after the system's 4–6GB overhead. The author lists recommendations ranging from the 5GB Llama 3.1 8B Abliterated to the 24GB Qwen 3.6 35B MoE Abliterated, with models runnable at the 16GB, 24GB, 32GB, and 48GB tiers. Image and video generation is provided by the app's in-app Create feature, including the Unfiltered group. Local runs don't send prompts or responses, only update queries and Cloud toggle counts. The Mac version remains in beta and has only been verified on a single M5 Pro 48GB.