Mac Compute Rentals: dedicated M3 Ultra Mac Studios with 512GB, by the 3, 7, or 14-day stretch.

What fits in 512GB of unified memory?

Weight sizes of large open models, taken from their published MLX conversions on Hugging Face and set against the 512GB in an M3 Ultra Mac Studio. These are sizes and memory arithmetic, not speed benchmarks.

Check availability & bookJump to the table

How to read this

512 GiB totalApple's "512GB" is binary gigabytes, so every size here is in GiB too.
Weights = filesThe sum of the .safetensors files in each repo, read from the Hugging Face API on Oct 14, 2026.
Left over512 minus the weights. macOS, your process and the KV cache (which grows with context length) all have to fit here.
Rule of thumbMemory arithmetic: params × bits ÷ 8. 400B params at 4-bit ≈ 200GB, plus a little for scales and unquantized layers.

Large open models that fit

ModelMLX quant (link)Weights (GiB)Left over (GiB)Verdict
gpt-oss-120b
117B total · Apache-2.0 · model card		MXFP4 + 8-bit59453Fits, lots of room
MiniMax-M2.1
~229B total · Custom license (see card) · model card		4-bit120392Fits, lots of room
Qwen3-235B-A22B-Instruct-2507
235B total / 22B active · Apache-2.0 · model card		4-bit123389Fits, lots of room
↳ same model8-bit233279Fits, lots of room
Qwen3.5-122B-A10B
~125B total / 10B active · Apache-2.0 · model card		8-bit122390Fits, lots of room
↳ same modelbf16 (unquantized)228284Fits, lots of room
DeepSeek-V4-Flash
284B backbone / 13B active · MIT · model card		4-bit141371Fits, lots of room
GLM-4.7
~358B total · MIT · model card		4-bit185327Fits, lots of room
↳ same model6-bit267245Fits, lots of room
Qwen3.5-397B-A17B
397B total / 17B active · Apache-2.0 · model card		4-bit208304Fits, lots of room
↳ same model6-bit301211Fits, lots of room
↳ same model8-bit393119Fits
Llama 4 Maverick 17B-128E
~402B total / 17B active · Llama 4 Community License · model card		4-bit210302Fits, lots of room
↳ same model6-bit304208Fits, lots of room
Llama 3.1 405B
405B dense · Llama 3.1 Community License · model card		4-bit215297Fits, lots of room
↳ same model8-bit402110Fits
MiniMax-M3
~428B total / ~23B active · MiniMax Community License · model card		4-bit225287Fits, lots of room
Qwen3-Coder-480B-A35B
480B total / 35B active · Apache-2.0 · model card		4-bit252260Fits, lots of room
DeepSeek-V3.2
~685B total · MIT · model card		4-bit352160Fits
Mistral Large 3
675B total / 41B active · Apache-2.0 · model card		4-bit360152Fits
GLM-5.2
~753B total · MIT · model card		MXFP4368144Fits
↳ same model4-bit390122Fits
GLM-5.3
~753B total · Custom license (see card) · model card		4-bit390122Fits
↳ same modelmixed 4/5-bit42092Fits, tight
Kimi K2.5
1T total / 32B active · Custom license (see card) · model card		3-bit41894Fits, tight
Kimi K2.6
1T total / 32B active · Custom license (see card) · model card		DQ3_K_M (≈3-bit mixed)43874Fits, tight

"Fits, lots of room" means at least 200 GiB left over, "Fits" means 100–199 GiB, and "Fits, tight" means under 100 GiB. Parameter counts come from each model card where stated, otherwise from Hugging Face's safetensors metadata (marked ~). Licenses vary, so read the model card before commercial use. Inclusion here doesn't mean we've tested the model on our machines.

What doesn't fit (or barely does)

Before you plan a big-model run

  1. Download time. The Studios have a ~200 Mbps business line. At full speed that's about 90 GB an hour, so a 400 GB model takes roughly 4.5 hours (400 × 8 ÷ 0.2 Gbps = 16,000 s). Real downloads are often slower. Budget for it, especially on a 3-day rental.
  2. Disk. Each Studio has a 2TB SSD shared with macOS, with at least 600GB free for your files (checked October 2026). Keep one copy of a 1T-class model (450–470 GB), not two.
  3. GPU wired memory. The mlx-lm README says that for models large relative to RAM you can raise the limit with sudo sysctl iogpu.wired_limit_mb=N, where N is larger than the model and smaller than RAM. You don't need to: both Studios already set iogpu.wired_limit_mb=480000 (about 469GB) at every boot, leaving the rest for macOS.
  4. Runtime support. The newest architectures (DeepSeek V4.x, Kimi K3, GLM-5.x) may need a recent mlx-lm or llama.cpp. Check before you book around one model.

Method and sources

Sizes: GET https://huggingface.co/api/models/<repo>?blobs=true, summing the sizes of the .safetensors files, on Oct 14, 2026. Each quant name in the table links to its repo, and each model name links to the original model card. Hardware: Apple's Mac Studio (2025) tech specs (M3 Ultra, up to 512GB unified memory, 819GB/s). Wired-memory note: the mlx-lm README, "Large Models". Repos change: re-check the sizes before you download, and tell us if a number here is out of date.

Already on the Studios

Some of these models are pre-cached, so you can skip the download:

Model (MLX)SizeStudio AStudio B
gpt-oss-120b, MXFP4 + 8-bit (mlx-community/gpt-oss-120b-MXFP4-Q8)59 GiBdownloadingdownloading
DeepSeek-V3.2, 4-bit (mlx-community/DeepSeek-V3.2-4bit)352 GiBdownloadingno
Llama 4 Maverick 17B-128E Instruct, 6-bit (mlx-community/Llama-4-Maverick-17B-128E-Instruct-6bit)304 GiBnodownloading

Shared models live read-only in /Users/Shared/models: a Hugging Face cache at /Users/Shared/models/hub, plus ready-to-use folders at /Users/Shared/models/by-name/<model> once a download finishes. No copy needed, and it doesn't use your disk space:

mlx_lm.generate --model /Users/Shared/models/by-name/gpt-oss-120b-MXFP4-Q8 --prompt "hello"
# or through the HF cache (read-only, so stay offline):
HF_HUB_CACHE=/Users/Shared/models/hub HF_HUB_OFFLINE=1 mlx_lm.generate --model mlx-community/gpt-oss-120b-MXFP4-Q8 --prompt "hello"

Download anything else into your own home folder (the default ~/.cache/huggingface).

Tools: Python, uv, mlx-lm (in a venv), llama.cpp, hf, git; Xcode on Studio B. See Support for details.

Run it on a whole 512GB Studio

Bare metal, one renter per machine, private access over Tailscale. 3 days $299 · 7 days $489 · 14 days $779.

Check availability & bookSee pricing