These are free AI models you run on your own computer, no internet needed. New to all this? Good, you are in the right place. We will explain every word, help you pick the one that fits your machine, and get you going in one command. The best ones can even see pictures.
You know chatbots like ChatGPT and Claude. They live on a company's computers far away, you pay every month, and everything you type gets sent over the internet to them. A local model is the same kind of smart helper, except it lives on your own computer and runs right there. Nothing leaves your machine.
Why would you want that? Four plain reasons. It is free, no monthly bill ever. It is private, your words and files never go anywhere. It works with the internet off, on a plane or in a cabin. And nobody caps how much you can use it. The trade is that it runs on your computer's memory, so a bigger, smarter model needs a bigger computer. That is the whole thing to understand, and this page helps you match a model to the machine you have.
Everything here is on our free Hugging Face page (Hugging Face is just the website where people share AI models, like an app store for them). Most of these are tuned-up versions of models built by Google, Meta, Alibaba, NVIDIA, Microsoft, Xiaomi and others. We made them run fast on Apple computers. We did not build the giants, and we thank the people who did down at the bottom.
Every AI page throws these words at you with no explanation. Here is each one in a single sentence. That is all they mean.
Vision model
A model that can look at a picture you give it, a screenshot, a photo, a camera frame, and answer questions about it, instead of only reading text.
Memory (RAM)
How much working space your computer has. On a Mac it is called unified memory. A model has to fit inside it, so this is the number that decides which models you can run.
GB
Gigabytes, the size of the download and roughly how much memory it needs while running. A 17 GB model wants a computer with more memory than 17 GB.
bf16 (full quality)
The model at full quality, nothing trimmed. These are the biggest files and the sharpest answers. Pick these when your computer has the memory to spare.
8-bit (about half the size)
The same model gently compressed to roughly half the size, still almost as good. A great middle choice for a medium computer.
4-bit (about a quarter the size)
Compressed further to about a quarter of the size. A little less sharp, but it lets a small or older computer run a model it otherwise could not.
mmproj
On some setups a vision model needs a small extra "eyes" file called mmproj. Ours bake the eyes right into the model, so you never have to think about it.
Abliterated
A build where the model's habit of refusing has been turned off, so it answers instead of saying no. Handy when you are building something unusual.
You have two ways to do this. The easy way picks and installs one for you. The hands-on way lets you grab any model from the table by clicking its link.
Want to pick yourself? Find your computer's memory in the first column (on a Mac, click the apple menu, then About This Mac, and read the Memory line), then grab one of these. Click a model name to go straight to its download page.
| Your computer's memory | A model that can see (vision) | A model for writing code and text |
|---|---|---|
| 8 to 16 GB most laptops |
Now a small computer can see too:AesCode-8B 4-bit 5.8 GB, sees and codes or MiMo-V2.6-9B 4-bit 6.0 GB or tiny Vision-Narrator-0.8B 0.65 GB | The friendly first model:Huihui Qwen3 8B 4-bit, 4.6 GB or Hermes 4 14B 4-bit, 8.3 GB |
| 24 to 32 GB a mid-range Mac or PC |
Step up the quality that can see:AesCode-8B 8-bit 9.9 GB or AesCode-8B bf16 17.5 GB, full quality, same for MiMo-9B 8-bit 10.5 GB, or the big AesCode-32B 4-bit 19.6 GB | A strong daily driver:Gemma 4 31B 4-bit, 17.3 GB or Qwen 3.6 27B 4-bit, 15.2 GB |
| 48 to 64 GB a high-memory Mac |
A bigger, sharper eye:AesCode-32B 8-bit 36.0 GB, the vision coder, nearly full quality or Muse-Glimmer-30B bf16, 59.6 GB | More room, more depth:Gemma 4 31B (huihui) 4-bit, 17.3 GB, with memory to spare for long work |
| 96 to 128 GB a maxed-out Mac |
The best, at full quality:AesCode-32B bf16 66.7 GB or NVIDIA Nemotron 74B 8-bit, 80.9 GB | The deepest thinker we carry:Llama 3.3 70B 8-bit, 75 GB |
Give these a picture and a question. A screenshot, a product photo, a camera frame, a page of a PDF. They look and they answer. These are the ones worth building around, and the reason we call this the vision king lineup. Each card tells you the size, the computer it wants, and a link to grab it.
Point it at a screenshot of a broken layout or a real web page and let it reason about what it sees.
Microsoft AesCode, fine-tuned on Alibaba's Qwen3-VL-32B-Instruct
Our flagship vision coder. It actually looks at the page or screenshot you give it and writes code against what it sees. Three sizes, so a mid-range computer can run the 4-bit and a maxed Mac can run full quality. Pick the biggest one your memory can hold.
bf16 · 8-bit · 4-bitAlibaba Qwen3-VL-32B, abliterated by huihui-ai
The same strong Qwen3-VL seeing and reasoning, squeezed to 4-bit so it fits a 32 GB Mac. This is the practical workhorse for screenshot and document work when you do not have a maxed machine.
huggingface.co/divinetribe/Huihui-Qwen3-VL-32B-Instruct-abliterated-4bit-mlxNewer and lighter, these fit a 24 to 32 GB machine, read screens and camera frames, and still write code.
Microsoft AesCode, fine-tuned on Alibaba's Qwen3-VL-8B-Instruct
The easiest vision coder to start with, and at 4-bit it is small enough for a 16 GB laptop. The little brother of AesCode-32B. Show it a screenshot of a broken page and it reads it and fixes the code.
bf16 · 8-bit · 4-bitXiaomi MiMo team, distilled from Alibaba's Qwen3.5-9B
A 9B agentic model with a vision tower. Small enough to feel quick, smart enough to run a tool loop, and it can look at what the agent is seeing. We fed it a live game camera frame and it read the overlay correctly. The 4-bit fits a 16 GB laptop.
bf16 · 8-bit · 4-bitNice Dreamz fine-tune of Alibaba's Qwen3.5-0.8B, trained on VizWiz
This is the little model inside our free RealTime AI Cam app for blind and low-vision people. Point a camera at a room or a letter, ask a question, and it answers out loud, all offline on the phone. Retrained on the real questions blind people asked about their own photos.
huggingface.co/divinetribe/Vision-Narrator-0.8B-4bit-mlxAbliterated so they answer instead of refusing, which matters when you are building something unusual and the model keeps saying no.
Converted from mlx-community/Muse-Glimmer-30B-bf16, refusal removed
One of our most downloaded vision builds. It does not just caption a picture, it reasons about fine detail, naming a specific car from its spoiler shape. Believed to be the first abliterated multimodal model running on Apple MLX.
huggingface.co/divinetribe/Muse-Glimmer-30B-Abliterated-MM-bf16NVIDIA Nemotron Omni, MLX conversion by mlx-community, refusal removed
A tri-modal model where text, vision and audio all work, with the refusal direction removed. Pick the size that fits your Mac. We believe it was the first abliterated Omni running in MLX. Three builds, so a 32 GB machine can run the 4-bit and a big machine can run full bf16.
bf16 · 8-bit · 4-bitNVIDIA Nemotron 3.5 Bolt, converted for MLX
The biggest vision model on the shelf, a mixture-of-experts design that only lights up about 9B of its weights per token, so it is smarter than its speed suggests. For the maxed-out Macs that want the heaviest seeing model we carry.
huggingface.co/divinetribe/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-MLX-8bitAesCode-32B with the refusal removed, on Alibaba's Qwen3-VL-32B
The uncensored twin of our flagship vision coder, for when the model keeps saying no to something you are trying to build. Same seeing and coding, with the refusal turned off. Three sizes so it fits any machine from a mid-range laptop up.
bf16 · 8-bit · 4-bitThese do not see pictures, they are for reading and writing. They are the daily drivers behind claude-code-local and they range from a 4.6 GB model that runs on almost anything to a 75 GB reasoner for a maxed machine.
The ones to keep loaded for everyday work.
Google DeepMind Gemma 4, refusal removed
Our most popular daily driver, with over 6,000 downloads. Google's Gemma 4 at a size that is comfortable on a 32 GB Mac and quick enough to feel instant. There is a huihui-ai abliterated build of the same model too if you prefer their refusal removal.
gemma-4-31b-it-abliterated-4bit-mlx · huihui buildAlibaba Qwen, refusal removed
A strong, fast Qwen that fits nicely on a 24 to 32 GB Mac. A great middle-of-the-road coder when Gemma is not to your taste.
huggingface.co/divinetribe/Qwen3.6-27B-abliterated-4bit-mlxLight models that leave room to breathe. Perfect for an older or base-model Mac.
Nous Research Hermes 4, on Alibaba Qwen3, refusal removed
A favorite, with over 5,500 downloads. Nous Research's Hermes 4 is a sharp, agreeable assistant that fits a 16 GB MacBook with room for your work. A fine first local model.
huggingface.co/divinetribe/Hermes-4-14B-abliterated-4bit-mlxGoogle DeepMind Gemma 4, refusal removed
The smaller Gemma 4, easy on a 16 GB Mac. There is also a text-only build of this one, trimmed down for pure chat and code when you do not need the extra parts.
gemma-4-12B-it-abliterated-4bit-mlx · text-only buildAlibaba Qwen3 8B, abliterated by huihui-ai
The smallest text model we carry. At 4.6 GB it runs on just about anything, including an 8 GB Mac. The one to start with if your machine is tight on memory.
huggingface.co/divinetribe/Huihui-Qwen3-8B-abliterated-v2-4bit-mlxWhen you have the memory and you want the most thoughtful answer.
Meta Llama 3.3 70B, refusal removed
Our most downloaded model of all, nearly 10,000 pulls. Meta's Llama 3.3 70B at 8-bit for the hardest reasoning you can run at home. This is the one for a 128 GB machine that wants depth over speed.
huggingface.co/divinetribe/Llama-3.3-70B-Instruct-abliterated-8bit-mlxIf you only read one thing, read step one. It does everything for you. The steps below it are for the curious who want to grab a specific model by hand.
Our free claude-code-local installer looks at your computer, picks a model that fits its memory, downloads it, and sets up a local AI you can run offline. It works on Mac, Linux and Windows. Copy the line for your computer, paste it into the Terminal app, and press return.
Every model above has a Hugging Face path. Download the one you want with the Hugging Face tool, swapping in the name from any card.
The vision models run with mlx-vlm, the Apple Silicon vision engine. Give it an image and a question.
The text and coding models run with mlx-lm, Apple's text inference library. Or just point the claude-code-local server at any model name and chat with it through Claude Code.
We want to be as clear and as grateful as we can about this. Almost everything on our shelf is a conversion or a fine-tune of someone else's brilliant work. We make these models run full and fast on Apple Silicon, but the giants were trained by the teams below, and the Apple Silicon inference itself is only possible because of the open-source libraries other people pour their lives into. Thank you, truly.
Several models here are abliterated, meaning the built-in refusal has been removed so the model answers instead of saying no. That technique rests on the research showing a single refusal direction can be ablated, and on the people who turned it into something anyone can use.
Everything here is free, open weight, and runs on the hardware you already own. Browse the whole library on Hugging Face, or let the one-command installer pick the right one for your Mac.