Skip to content

Tools / LLM GPU calculator / 96 GB Mac Studio

What LLMs can run on a 96 GB Mac Studio?

Which of 22 open LLMs fit one 96 GB Mac Studio, and how many it takes for the rest, at FP16, FP8 and 4-bit.

Size your own scenario

Short answer

One 96 GB Mac Studio (about 72 GB usable by the GPU) runs 17 of the 22 open LLMs sized here at 4-bit, 12 at FP8 and 9 at FP16, for one user at an 8K context. At 4-bit: every model except Llama 3.1 405B, Llama 4 Maverick, Qwen3-235B-A22B, DeepSeek-V3 / R1 and GLM-5.3. The largest that fits one Mac is gpt-oss-120b (116.8B parameters), at 4-bit. Bigger models split their layers across several Macs: Qwen3-235B-A22B takes 3 at 4-bit.

Updated . One user, 8K context, KV cache at FP16, plus 10% headroom for activations and the runtime.

Every model on the 96 GB Mac Studio

ModelFP16/BF16FP84-bit (INT4)
Llama 3.1 8B1 Mac19 GB1 Mac10 GB1 Mac6.1 GB
Qwen3 8B1 Mac20 GB1 Mac10 GB1 Mac6.3 GB
Qwen3 14B1 Mac34 GB1 Mac18 GB1 Mac11 GB
Phi-4 14B1 Mac35 GB1 Mac18 GB1 Mac11 GB
gpt-oss-20b1 Mac47 GB1 Mac23 GB1 Mac13 GB
LFM2.5-24B-A2B1 Mac54 GB1 Mac27 GB1 Mac15 GB
Mistral Small 3.2 24B1 Mac55 GB1 Mac28 GB1 Mac16 GB
Gemma 3 27B1 Mac62 GB1 Mac32 GB1 Mac18 GB
Qwen3-30B-A3B1 Mac69 GB1 Mac35 GB1 Mac20 GB
Qwen3 32B2 Macs75 GB1 Mac39 GB1 Mac22 GB
Qwen2.5 32B2 Macs75 GB1 Mac39 GB1 Mac22 GB
Mixtral 8x7B2 Macs105 GB1 Mac53 GB1 Mac30 GB
Llama 3.3 70B3 Macs160 GB2 Macs81 GB1 Mac46 GB
Qwen2.5 72B3 Macs165 GB2 Macs84 GB1 Mac47 GB
Kolibri-13 Macs174 GB2 Macs87 GB1 Mac48 GB
Llama 4 Scout4 Macs243 GB2 Macs122 GB1 Mac68 GB
gpt-oss-120b4 Macs260 GB2 Macs130 GB1 Mac72 GB
Qwen3-235B-A22B8 Macs524 GB4 Macs263 GB3 Macs145 GB
Llama 4 Maverick13 Macs894 GB7 Macs448 GB4 Macs247 GB
Llama 3.1 405B13 Macs907 GB7 Macs456 GB4 Macs253 GB
DeepSeek-V3 / R121 Macs1,492 GB11 Macs746 GB6 Macs411 GB
GLM-5.323 Macs1,654 GB12 Macs827 GB7 Macs455 GB

Counts are for a layer split, which works with any number of Macs. Below each count is the memory needed: (weights + KV cache) ÷ 0.9. Capacity only: speed depends on memory bandwidth and the serving stack.

Memory is step one. Throughput, latency, power and the network decide the real build.

Size my actual workload

96 GB Mac Studio specifications

Memory
96 GB unified, about 72 GB usable by the GPU
Memory bandwidth
819 GB/s
Typical system
Desktop
Counted as
Layer split across cards (llama.cpp style)

Questions about the 96 GB Mac Studio

What LLMs can run on a 96 GB Mac Studio?

For one user at an 8K context, these fit one 96 GB Mac Studio. At 4-bit: every model except Llama 3.1 405B, Llama 4 Maverick, Qwen3-235B-A22B, DeepSeek-V3 / R1 and GLM-5.3. At FP8: every model except Llama 3.3 70B, Llama 3.1 405B, Llama 4 Scout, Llama 4 Maverick, Qwen3-235B-A22B, Qwen2.5 72B, DeepSeek-V3 / R1, gpt-oss-120b, GLM-5.3 and Kolibri-1. At FP16/BF16: Llama 3.1 8B, Qwen3 8B, Qwen3 14B, Phi-4 14B, gpt-oss-20b, LFM2.5-24B-A2B, Mistral Small 3.2 24B, Gemma 3 27B and Qwen3-30B-A3B. That counts the 72 GB of unified memory macOS lets the GPU use by default.

What is the largest LLM a 96 GB Mac Studio can run?

Of the models sized here, for one user at an 8K context: at 4-bit, gpt-oss-120b (116.8B parameters, 72 GB); at FP8, Mixtral 8x7B (46.7B parameters, 53 GB); at FP16, Qwen3-30B-A3B (30.5B parameters, 69 GB). Mixture-of-experts models count every expert, because all of them must be in memory.

How many 96 GB Mac Studios does Llama 3.3 70B need?

For one user at an 8K context it needs 160 GB at FP16/BF16, 81 GB at FP8 and 46 GB at 4-bit: 3 Macs at FP16, 2 Macs at FP8 and one Mac at 4-bit, with a layer split.

Can a 96 GB Mac Studio run DeepSeek-V3 / R1?

Not on one Mac: with all 671B parameters resident it needs 746 GB at FP8 and 411 GB at 4-bit for one user at an 8K context. That is 11 Macs at FP8 and 6 Macs at 4-bit, with a layer split.

Can I combine several 96 GB Mac Studios?

Yes, with a layer split: each Mac holds some of the model's layers, as llama.cpp does. Memory adds up across them; speed does not, because they take turns on each token. The counts on this page assume a layer split.

Other GPUs

Different context length, more users, or a model that isn’t listed? The LLM GPU calculator sizes any Hugging Face model for any of these GPUs.

Need to deploy private AI?

We design it, validate it, source it, deploy it and help run it.

Get my 48-hour blueprint

Give us your workload, users, budget and timeline. Within 48 hours we'll give you a validated cloud/on-prem architecture, BOM, expected costs and sourcing options.