Which local LLM fits your Mac? A guide by memory size
The useful question is rarely "which model is best" but "which model will run well on this Mac". Memory decides most of the answer. This guide explains why, then gives concrete picks for 8, 16, 24 and 32 GB and more.
Why memory decides what you can run
A language model is a very large set of numbers, called weights. To write each word, or piece of a word, the computer works through those weights. So the whole model has to sit in memory while it answers. If it does not fit, the Mac starts moving data back and forth to the disk, and answers slow to a crawl.
On Apple Silicon Macs, the processor and the graphics cores share one pool of memory, called unified memory. That suits local models well, because a model can use a large part of your RAM. It also means the model competes with everything else you have open: macOS, your browser, your design tools.
A model also needs working memory on top of its file, to hold the conversation and the text you give it. That is why the memory a model needs is always higher than its download size.
To check how much memory your Mac has, open the Apple menu and choose About This Mac. The local model library in Pandore also shows your processor and memory at the top.
Quantization, in plain words
Each weight is a number. Models are usually published with each number stored at high precision, 16 bits. At that precision, a model with 8 billion parameters needs about 16 GB just for its weights: 8 billion times 2 bytes.
Quantization stores each number with fewer bits. Think of saving a photo as a smaller JPEG: you lose a little detail, and the file shrinks a lot. Most files in Pandore's catalog use a 4-bit scheme called Q4_K_M, which keeps some sensitive parts of the model at higher precision. With it, Qwen 3 8B is a 5.0 GB download instead of about 16 GB.
A rule of thumb from the catalog: at 4 bits, each billion parameters costs roughly 0.6 GB of download. The 14B models weigh 8 to 9 GB; Gemma 3 27B weighs 16.5 GB. The smallest model, SmolLM2 360M, uses 8-bit quantization (Q8_0), which keeps more precision, and is still under 400 MB.
You do not have to choose a quantization in Pandore. Each model in the library is a single file with its quantization already picked, and the card shows which one.
Three kinds of models
The library sorts its 43 models into three categories, and you can filter by each one.
- General models handle everyday text: summaries, rewrites, drafts, translations, questions about a document. Gemma 3, Llama 3.2, Qwen 2.5, Ministral 3, Mistral Nemo, Granite 3.3, OLMo 2, SmolLM2 and the smaller LFM2.5 models are here.
- Coding models have been trained with a focus on code. In the catalog, they are the five Qwen 2.5 Coder models, from 0.5B to 14B.
- Reasoning models work through a problem step by step before they answer. They do better on plans, comparisons and logic, and they are slower, because they write more before the answer. This group holds Qwen 3 (from 1.7B up), Phi-3 and Phi-4, DeepSeek R1 Distill, SmolLM3, LFM2.5 8B A1B and the Ministral 3 Reasoning models.
Picks by memory size
Every model in the catalog has a recommended amount of memory, at one of four levels: 4, 8, 16 or 32 GB. The library's For this Mac filter, selected by default, shows only the models whose recommendation fits your RAM. Each card also says "Good fit" or "May be slow" for your machine.
Here are concrete starting points. Sizes in parentheses are download sizes, rounded.
| Your Mac's memory | General | Coding | Reasoning |
|---|---|---|---|
| 8 GB | Gemma 3 4B (2.5 GB) | Qwen 2.5 Coder 3B (2.1 GB) | Qwen 3 4B (2.5 GB) |
| 16 GB | Ministral 3 8B (5.2 GB) | Qwen 2.5 Coder 7B (4.7 GB) | Qwen 3 8B (5.0 GB) |
| 24 GB | Ministral 3 14B (8.2 GB) | Qwen 2.5 Coder 14B (9.0 GB) | Qwen 3 14B (9.0 GB) |
| 32 GB and more | Gemma 3 27B (16.5 GB) | Qwen 2.5 Coder 14B (9.0 GB) | Qwen 3 30B A3B (18.6 GB) |
8 GB
Every model recommended for 4 or 8 GB fits: 24 models, none larger than 2.5 GB. Start with Gemma 3 4B or Qwen 3 4B. Phi-4 Mini (2.5 GB) and Llama 3.2 3B (2.0 GB) are good alternatives. For the fastest replies, try Qwen 2.5 1.5B (1.1 GB) or Gemma 3 1B (0.8 GB). On 8 GB, the model shares memory with everything else, so close heavy apps if answers feel slow.
16 GB
This adds 17 models recommended for 16 GB. The 7B and 8B models, from 4.5 to 5.2 GB, are the comfortable middle: Ministral 3 8B, Qwen 3 8B, Granite 3.3 8B, Qwen 2.5 Coder 7B and DeepSeek R1 Distill 7B. The 12B to 14B models, from 7.3 to 9.0 GB, also count as a fit, but they leave less room for the rest of your work. Gemma 3 12B (7.3 GB) is a reasonable step up; see how your Mac feels with your usual apps open.
24 GB
The library recommends the same list as at 16 GB, because the two largest models ask for 32 GB. What changes is headroom: the 14B models become comfortable. Qwen 3 14B, Phi-4 14B (8.4 GB), Ministral 3 14B and Qwen 2.5 Coder 14B are the ones to try. Other amounts between 16 and 32 GB follow the same logic.
32 GB and more
Every model in the catalog fits, including the two largest: Gemma 3 27B (16.5 GB) and Qwen 3 30B A3B (18.6 GB). For code, Qwen 2.5 Coder 14B stays the pick: it is the largest coding model in the catalog.
Speed or quality
Within a family, larger models generally follow instructions more closely, write more naturally and make fewer mistakes. They are also slower, because the Mac works through more data for every word, and they take longer to load.
Two details change that trade-off:
- Mixture-of-experts models. Qwen 3 30B A3B has 30 billion parameters but uses about 3 billion for each word it writes. LFM2.5 8B A1B uses about 1 billion out of 8. They tend to answer faster than a classic model of the same size, but the whole file still has to fit in memory.
- Reasoning models take longer before the answer, because they think on the page first. Use them when the question deserves it, not for a quick rewrite.
A practical setup is one small model and one larger model. In Pandore, the library's Auto mode then chooses among your downloaded models for each request: it leans toward a coding model when the request is about code, a reasoning model for analysis, and a fast model for short rewrites. The model stays loaded while Pandore is open, so you do not wait for it to start each time, but it also keeps using its share of memory.
What local models do well, and where they fall short
Local models are good at work on text you give them:
- Summaries, rewrites, translations and extracting facts from a document.
- Private material. The text stays on your Mac.
- Offline work. Once downloaded, a model needs no internet connection, no API account and no usage bill.
They fall short in a few predictable places:
- Facts from memory. Smaller models remember less, and no model knows what happened after its training. Give them the source text instead of asking them to recall it.
- Numbers. Like any language model, they can get arithmetic wrong. For anything priced, read how to write a client proposal without invented numbers.
- Work that needs tools. In Pandore, local models answer in chat and work on text. They do not edit files in your project folder, and Missions run with remote agents such as Claude Code.
The simplest way to get both is Settings › AI in Both mode: Pandore keeps simple or sensitive requests local and sends tasks that need tools or more capacity to your remote agent. In that mode, a message with an image goes to the remote agent. For more on that choice, see local vs cloud AI for private data, and for the first download, the step-by-step guide to running a local LLM on your Mac.
Each card also shows the model's license. Most are Apache 2.0; others carry their own terms, such as Gemma Terms, Llama 3.2, Qwen Research or the LFM Open License. Read it before commercial use.
