How to run DeepSeek locally on a Mac

You typed "run DeepSeek locally Mac" because you want the model on your own machine, not behind someone's login page. Good news: it works, and it's free. DeepSeek's weights are open, so on an Apple Silicon Mac you download a quantized copy and run it offline. The one catch is size — the full models are enormous, so the move is to pick a quantized build that fits your RAM. Here's the honest version, including a table so you know which size your Mac can actually hold.

What "DeepSeek" actually is

DeepSeek isn't one model. It's a family of open-weight LLMs (the V3 and R1 lines), known for strong reasoning and coding. The headline models are very large — hundreds of billions of parameters — which is why the popular way to run them on a laptop is a distill or a quantized build: a smaller version, or the same weights compressed to fewer bits so the file shrinks and fits in less memory. On a Mac the format you want is GGUF, the file type the local runners read. Everything below assumes Apple Silicon (M1 or newer); Intel Macs technically run these but slowly enough that it's not worth it.

The two simplest paths

You have two good front doors, both free. Pick by whether you like a terminal.

Path A — Ollama (one command)

1. Download Ollama from its site and install it (it ships a tiny menu-bar app and a CLI).
2. Open Terminal and run ollama run deepseek-r1:8b. Ollama pulls the quantized GGUF, loads it, and drops you into a chat prompt.
3. To try a different size, swap the tag — for example deepseek-r1:14b — as long as it fits your RAM (see the table below).
4. That's it. It runs offline from here; turn off Wi-Fi and it still answers.

Path B — LM Studio (a GUI, no terminal)

1. Download LM Studio and open it.
2. Use the built-in model browser to search "DeepSeek," then pick a quantized GGUF whose size fits your Mac. LM Studio shows an estimate of whether a build will fit before you download.
3. Hit load, then chat in the window. LM Studio can also expose a local OpenAI-compatible server if you want to point your own code at it.

Jan and GPT4All are two more free, open-source GUIs that do the same job if you prefer them. All of these are RAM-bound: the model has to fit in your Mac's unified memory, so the size you choose is the whole game.

Which DeepSeek size fits your Mac?

The single rule of local LLMs: the model must fit in memory, with a few gigabytes left for macOS. On Apple Silicon, RAM and VRAM are the same unified pool, which is why a 16 GB Mac is a real constraint, not a suggestion. Rough expectations, by quantized size:

These are rough bands, not promises — exact file sizes vary by the specific build and quant level you pick. The pattern holds though: a quantized 7–8B fits comfortably on 16 GB, and bigger sizes want 32–64 GB or more. If you're not sure how much headroom you have, our RAM guide for local AI walks through the math.

The honest trade-offs

Running locally costs you two things. First, speed: a 14B or 32B on an M-series chip generates at a readable pace but won't match a cloud datacenter, and the bigger the model the slower the tokens. Second, ceiling: a quantized distill is smaller than the full DeepSeek, so on the hardest problems it gives up some quality. What you get in exchange is everything that matters about "local" — it's private, it runs with Wi-Fi off, there's no meter, and nobody can deprecate or reprice the file sitting on your disk.

Where Outlier fits

If the terminal-and-GGUF route sounds like more setup than you want, that's the gap Outlier fills: one signed Mac download, no account, no Docker, no command line. Outlier doesn't ship DeepSeek by name. Its own models are built on the Qwen family — another strong open-weight line — and cover the same daily reasoning and coding work most people open DeepSeek for. On a 54-prompt comparison, Outlier's Core 27B matched Claude Opus on 98.9% of rubric checks (the full results are public). And because of a patent-pending paged inference engine, Outlier can run a model bigger than the Mac's RAM — a 397B-parameter model at roughly 11 GB peak memory on a 64 GB Mac Studio — which the RAM-bound GGUF tools above can't do.

If you specifically want DeepSeek itself, Outlier supports bring-your-own-model: import an MLX build and run it inside the same app. So the choice isn't DeepSeek or Outlier — it's a free GGUF runner when you want that exact model, or a no-setup Mac app for the everyday work, with a path to load your own model either way. If you're choosing a runner first, our Outlier vs Ollama breakdown lays it out plainly.

Frequently asked questions