Mac Studio local LLM memory: what each size runs
Apple lists 36 GB as the smallest Mac Studio and 512 GB as the largest. At the default macOS GPU share, the smallest Mac Studio gives the GPU about 27 GiB and the largest gives about 384 GiB. On the largest configuration, the biggest model in our list that runs is Qwen3-235B-A22B at Q8_0 (234.4 GiB–239.5 GiB at 8,192 tokens). On the smallest, it is Qwen2.5-32B-Instruct at Q4_K_M.
Chips and memory sizes Apple lists
| Chip | GPU | Memory bandwidth | Unified memory |
|---|---|---|---|
| Apple M5 Max (18-core CPU) | 32-core GPU, or 40-core GPU as an option | 460 GB/s (614 GB/s with the 40-core GPU) | 36 GB, 48 GB, 64 GB, 128 GB (36 GB is the base; 48, 64 and 128 GB go with the 40-core GPU configuration.) |
| Apple M5 Ultra (30-core CPU) | 64-core GPU, or 80-core GPU (36-core CPU) as an option | 1.2 TB/s | 96 GB, 256 GB, 512 GB (96 GB is the base; 256 and 512 GB go with the 80-core GPU configuration.) |
From Apple's Mac Studio technical specifications, read 2026-10-18. Check that page before buying: Apple ties some memory sizes to a specific chip configuration. Bandwidth is Apple's figure. This site does not turn it into tokens per second and has no measurement on these machines.
What each memory size runs
| Unified memory | GPU budget | Models that run | Tight | Largest that runs |
|---|---|---|---|---|
| 36 GB | 27.0 GiB | 16 of 22 | 0 | Qwen2.5-32B-Instruct, Q4_K_M |
| 48 GB | 36.0 GiB | 17 of 22 | 0 | Mixtral-8x7B-Instruct-v0.1, Q4_K_M |
| 64 GB | 48.0 GiB | 17 of 22 | 3 | Mixtral-8x7B-Instruct-v0.1, Q6_K |
| 96 GB | 72.0 GiB | 20 of 22 | 1 | Qwen3-Next-80B-A3B-Instruct, Q6_K |
| 128 GB | 96.0 GiB | 21 of 22 | 0 | GLM-4.5-Air, Q5_K_M |
| 256 GB | 192.0 GiB | 22 of 22 | 0 | Qwen3-235B-A22B, Q5_K_M |
| 512 GB | 384.0 GiB | 22 of 22 | 0 | Qwen3-235B-A22B, Q8_0 |
Budget is the default share macOS gives the GPU: 67% up to 32 GB and 75% above (the rule and sources are on each tier page). Verdicts use the catalog's llama.cpp GGUF files, one sequence, fp16 KV cache and 8,192 tokens, over the 22 models in the Mac list. Mixture-of-experts models count at full file size. Sizes without a tier page (36, 256 and 512 GB) use the same rule. These are memory estimates, not runs on a Mac.
Data sources
- Apple Mac Studio technical specifications — Chip, GPU core, bandwidth and memory options read 2026-10-18.
- Apple silicon unified memory and LLM sizing (LLMS-Explorer) — Quotes Apple's Metal Compute talk: 32 GB Mac gives the GPU 21 GB, 64 GB gives 48 GB. Newer 128 GB M4 Max logs show about 84%, so the share is not a constant.
- recommendedMaxWorkingSetSize (Apple Developer Forums) — Discussion of the default GPU working-set limit on Apple silicon.
- Model file sizes and architectures — catalog dataset 2026-10-02.1; GGUF bytes come from the Hugging Face repos listed on each model page