Mac Compute Rentals › Guides

How to run DeepSeek-V3.2 on a 512GB Mac Studio (M3 Ultra) with MLX

DeepSeek-V3.2 is a 671B-parameter mixture-of-experts model. Its 4-bit MLX conversion, mlx-community/DeepSeek-V3.2-4bit, is 352 GiB of weights (summed from the Hugging Face API, Oct 14, 2026). That's too big for any single NVIDIA GPU and for every 256GB machine, but it fits in one M3 Ultra Mac Studio with 512GB of unified memory with plenty of room left for the KV cache.

Memory math

On a stock Mac, macOS caps GPU-wired memory well below total RAM, and the mlx-lm README recommends raising it with sudo sysctl iogpu.wired_limit_mb=N for large models. Our Studios already set it at every boot, so you don't need sudo.

1. Install mlx-lm

Use a recent mlx-lm; new architectures land there first.

uv venv ~/mlx && source ~/mlx/bin/activate
uv pip install -U mlx-lm

2. Get the weights

On our Studio A, DeepSeek-V3.2 4-bit is being pre-cached in the shared read-only model folder. Once it's there, skip the download:

mlx_lm.generate --model /Users/Shared/models/by-name/DeepSeek-V3.2-4bit --prompt "hello"

Otherwise download it to your home folder. At ~200 Mbps, 352 GiB takes about 4 hours at full line speed.

hf download mlx-community/DeepSeek-V3.2-4bit

3. Chat or serve it

mlx_lm.chat --model mlx-community/DeepSeek-V3.2-4bit
# OpenAI-compatible server on port 8080 (reachable over your Tailscale share)
mlx_lm.server --model mlx-community/DeepSeek-V3.2-4bit --host 0.0.0.0 --port 8080

Tips

• Keep one copy of the weights on disk; a second copy of a 350 GiB model eats most of the free SSD.
• Long contexts grow the KV cache. If you hit memory pressure, shorten the context or try a smaller quant.
• Compare against other big models on our what fits in 512GB table.