What LLM can a Mac with 96 GB unified memory run?
96 GB of unified memory is not 96 GB of VRAM. macOS lets the GPU wire about 75% by default, so this page budgets 72.0 GiB for weights, KV cache and runtime. The largest model in our list that fits is Qwen3-Next-80B-A3B-Instruct at Q6_K (61.9 GiB–63.6 GiB at 8,192 tokens). 20 of 22 models run, 1 are tight, 1 do not fit.
Model by model on 96 GB
| Model | Result | Quant to use | Estimated total | Max context |
|---|---|---|---|---|
| Phi-4-mini-instruct | Runs | Q8_0 | 4.9 GiB–5.5 GiB | 131,072 |
| Qwen3-4B | Runs | Q8_0 | 5.2 GiB–5.8 GiB | 40,960 |
| Mistral-7B-Instruct-v0.3 | Runs | Q8_0 | 8.3 GiB–8.9 GiB | 32,768 |
| Qwen2.5-7B-Instruct | Runs | Q8_0 | 8.1 GiB–8.7 GiB | 32,768 |
| DeepSeek-R1-Distill-Llama-8B | Runs | Q8_0 | 9.1 GiB–9.7 GiB | 131,072 |
| Qwen3-8B | Runs | Q8_0 | 9.4 GiB–10.0 GiB | 40,960 |
| GLM-4-9B-0414 | Runs | Q8_0 | 9.8 GiB–10.3 GiB | 32,768 |
| Mistral-Nemo-Instruct-2407 | Runs | Q8_0 | 13.5 GiB–14.2 GiB | 131,072 |
| phi-4 | Runs | Q8_0 | 16.2 GiB–17.0 GiB | 16,384 |
| Qwen3-14B | Runs | Q8_0 | 16.0 GiB–16.8 GiB | 40,960 |
| Qwen2.5-14B-Instruct | Runs | Q8_0 | 16.2 GiB–17.0 GiB | 32,768 |
| Mistral-Small-24B-Instruct-2501 | Runs | Q8_0 | 24.7 GiB–25.6 GiB | 32,768 |
| Qwen3-30B-A3B | Runs | Q8_0 | 31.1 GiB–32.2 GiB | 40,960 |
| Qwen3-32B | Runs | Q8_0 | 34.6 GiB–35.7 GiB | 40,960 |
| Qwen2.5-32B-Instruct | Runs | Q8_0 | 34.6 GiB–35.7 GiB | 32,768 |
| DeepSeek-R1-Distill-Qwen-32B | Runs | Q8_0 | 34.6 GiB–35.7 GiB | 121,361 |
| Mixtral-8x7B-Instruct-v0.1 | Runs | Q8_0 | 47.3 GiB–48.7 GiB | 32,768 |
| Llama-3.1-Nemotron-70B-Instruct-HF | Runs | Q6_K | 56.5 GiB–58.2 GiB | 28,508 |
| Qwen2.5-72B-Instruct | Runs | Q6_K | 58.4 GiB–60.0 GiB | 22,769 |
| Qwen3-Next-80B-A3B-Instruct | Runs | Q6_K | 61.9 GiB–63.6 GiB | 20,939 |
| GLM-4.5-Air | Tight | Q4_K_M | 69.5 GiB–71.4 GiB | 22,288 |
| Qwen3-235B-A22B | No | Smallest file is 132.4 GiB | n/a | n/a |
llama.cpp GGUF files, one sequence, fp16 KV cache, 8,192 tokens. Mixture-of-experts models keep all weights resident, so their file size sets the memory need, not their active parameters. Estimates, not measurements on a Mac.
Where the 72 GiB comes from
Apple does not publish a per-size rule. Its Metal talk gives 21 GB for the GPU on a 32 GB machine and 48 GB on a 64 GB one, which is two thirds and three quarters. This page uses two thirds up to 32 GB and three quarters above that. Some newer 128 GB machines report about 84%, so on those you get more than this page assumes. The limit can be raised with sudo sysctl iogpu.wired_limit_mb=… on recent macOS; raising it leaves less for the system.
The catalog's own Apple records use a different reserve: M2 Max 64GB reserves 10 GiB and M4 Max 128GB reserves 12 GiB, which assumes you have raised the limit. Their numbers will be higher than the 64 and 128 GB rows here.
Data sources
- 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