Learning path · Local models

Hugging Face for beginners: find, size and download local AI models

Almost every open AI model you can run at home is published on Hugging Face. This guide shows you how to find the right one, check that you're allowed to use it, pick a file that fits your computer, and load it into Ollama or LM Studio.

Checked October 15, 2026 against the Hugging Face Hub docs, the Ollama docs and the LM Studio docs. Text-only guide: button names are quoted exactly as they appear on each site.

On this page: What Hugging Face is · Find a model · Read a model card · Licenses · GGUF and quantization · Pick a size · Download into Ollama · Download into LM Studio · Fixing common problems

What Hugging Face is

Hugging Face is the main public library for AI models, a bit like an app store crossed with GitHub. Companies such as Meta, Google, Microsoft, Mistral, Alibaba (Qwen) and OpenAI publish their open-weight models there, and the community uploads ready-to-run versions of them. Browsing and downloading most models is free and needs no account. You only need a free account for "gated" models, where the publisher asks you to accept a license first.

Models are the downloadable model files. This is what you'll use.
Datasets are training data. You can ignore them.
Spaces are free online demos. Useful for trying a model in your browser before you download it.

How to find a model

Go to huggingface.co/models. You'll see a long list of models with a filter panel on the left.
Under Tasks, pick what you want: "Text Generation" for chat and coding, "Text-to-Image", "Text-to-Video" or "Text-to-Speech".
Under Libraries, tick GGUF if you want a chat or coding model for Ollama or LM Studio. On a Mac, MLX versions also work in LM Studio.
Use the Sort menu (top right of the list) and choose "Most downloads" or "Trending". Popular models are usually the better-tested ones.
Or skip the searching: our model picker lists checked models for each hardware tier, each linked to its official page.

Official vs community uploads. The part before the slash is the publisher, for example Qwen/Qwen3-8B is from the Qwen team. Ready-to-run GGUF copies are usually made by well-known community uploaders (for example bartowski or unsloth) or by the publisher itself. Prefer the official page for facts and license, and a popular, recently updated uploader for GGUF files.

Tip: tell Hugging Face your hardware. Signed in, open your account settings, go to Local Apps, and add your hardware (GPU or Mac). Model pages then show which file sizes should fit your machine (Hugging Face: Local Apps).

How to read a model card

Every model page opens on its Model card tab, written by the publisher (what model cards contain). Look for five things:

Size (parameters). Shown as "7B", "32B" and so on, in the name or in the "Model size" box on the right. B means billion. Bigger is smarter but needs more memory.
Active parameters. Names like "30B-A3B" mean a mixture-of-experts model: 30B in total, but only 3B work on each word, so it runs fast. It still needs memory for all 30B.
License. A tag near the top, such as apache-2.0, mit or other. See the next section.
Hardware notes. Many image, video and voice cards state the GPU memory they need. Trust these over guesses.
Files and versions. This tab lists the actual files and their sizes. That size is roughly how much memory you need.

Licenses: can you use it for work?

Apache-2.0 and MIT: use it for anything, including business, as long as you keep the license notice.
Community licenses (Llama, Gemma, Stability AI): generally allowed for business, with conditions. Stability AI's, for example, is free under $1M yearly revenue.
Non-commercial (CC-BY-NC, "Non-Production", FLUX.1 [dev]): personal, research or testing only. Don't use them in paid work without buying a license.
"other": read the license file in the model's Files tab. Our model picker shows each model's license in its own column.

A community GGUF copy keeps the original model's license. Always check the original.

Gated models. Some pages show a box asking you to share your contact details and accept terms. Sign in, read the terms, and click the button in that box (for example "Agree and access repository"). Access is often instant; some publishers review requests by hand (Hugging Face: gated models).

What GGUF and quantization mean

GGUF is a single-file format for running language models on ordinary computers. It's what Ollama, LM Studio and llama.cpp use (Hugging Face: GGUF). Quantization shrinks a model by storing its numbers with less precision. One model is usually offered in many quantized files, named like this:

Label	Size vs original	Quality	When to pick it
Q8_0	about half	Practically identical	You have plenty of memory
Q6_K	about 40%	Very close	A safe upgrade if it fits
Q4_K_M	about 30%	Small, usually unnoticeable loss	The default. Start here.
Q3_K_M / Q2_K	about 20–25%	Noticeably worse	Only to squeeze a big model in

Image and video models use other formats (.safetensors, FP8 or GGUF copies for ComfyUI), but the same idea applies: smaller files trade a little quality for fitting in less memory.

How to pick a size for your RAM or VRAM

The rule: the file size, plus 2–4GB for the conversation, must fit in your GPU memory (VRAM), or in about 70% of a Mac's memory. A rough guide at Q4_K_M:

Your computer	Model size that runs well	Example
8GB VRAM, or a 16GB Mac or PC	up to about 8B (file ≈ 5GB)	Qwen3 8B, Qwen2.5-Coder 7B
12–16GB VRAM, or a 32GB Mac	up to about 14B, or 20–30B mixture-of-experts	gpt-oss-20b, Qwen3-Coder 30B-A3B (tight)
24GB VRAM, or a 48–64GB Mac	up to about 32B (file ≈ 20GB)	Qwen2.5-Coder 32B, Devstral Small
48GB+ VRAM, or a 96–128GB Mac	70B and large mixture-of-experts	Llama 3.3 70B, gpt-oss-120b

To find your memory: on Windows, open Task Manager → Performance → GPU and read "Dedicated GPU memory". On a Mac, Apple menu → About This Mac and read "Memory". Full per-model picks are in the model picker.

Download a Hugging Face model into Ollama

Install Ollama first (steps in Run AI models on your own computer). Then:

On Hugging Face, open a GGUF model page, for example bartowski/Qwen2.5-Coder-7B-Instruct-GGUF.
Click Use this model (top right) and choose Ollama. Pick a quantization from the list, for example Q4_K_M, and copy the command it shows.
Open Terminal (Mac) or PowerShell (Windows), paste the command and press Enter. It looks like:
CommandPaste this into Terminal (Mac) or PowerShell (Windows).
```
ollama run hf.co/bartowski/Qwen2.5-Coder-7B-Instruct-GGUF:Q4_K_M
```
What you should see: "pulling manifest", then a progress bar while the file downloads, then "success" and a >>> prompt. Type a question to test it. Type /bye to exit.
From now on the model appears in ollama list and in any app connected to Ollama, under the same hf.co/... name.

This works for any public GGUF model on Hugging Face (Hugging Face: use Ollama with any GGUF model). Many popular models are also in Ollama's own library under short names like qwen3:8b.

Download a Hugging Face model into LM Studio

Open LM Studio and click the Discover tab (the magnifying glass in the left sidebar).
Type the model name, for example "Qwen2.5 Coder 7B". Results come straight from Hugging Face.
Click a result. On the right, LM Studio lists the available files and marks the ones likely to fit your machine, and warns when one is probably too large.
Pick a Q4_K_M file (or an MLX 4-bit file on a Mac) and click Download. A progress bar appears at the bottom.
When it finishes, go to the Chat tab, click the model selector at the top, choose the model and start typing.

Shortcut: on a Hugging Face model page, Use this model → LM Studio opens the model directly in LM Studio (LM Studio: download a model).

Fixing common problems

"Use this model" has no Ollama or LM Studio option. The page isn't a GGUF model. Search for the model name plus "GGUF" to find a converted copy.
Error 401 or "access denied". The model is gated. Sign in on Hugging Face and accept its terms first. Or use an ungated community GGUF copy, which keeps the same license.
Ollama says the model requires more system memory, or it's extremely slow. The file is too big. Pick a smaller quantization (Q4 instead of Q8) or a smaller model. Run ollama ps: anything less than "100% GPU" means part of it spilled into slower memory.
LM Studio says "Failed to load model". Usually not enough memory. Choose a smaller file, close other apps, or lower the context length in the load settings.
The download stalls. Large files can take a long time. Check free disk space (models are 2–70GB) and restart the download; both apps resume.
The model rambles or answers in strange formats. Use the "Instruct" or "it" (instruction-tuned) version, not the "Base" version.