TechCompare LogoTechCompare
Showing centimeters. Press to switch to inches.Showing inches. Press to switch to centimeters.

Which LLMs Fit on a 512 GB Mac?

By TechCompare · Updated · How we check

The calculator gives a 512 GB Mac 504 GB for models, after 8 GB for macOS and your apps. 89 of the 108 models fit there at Q4_K_M with 8K context. Step-5-Preview 604B (MoE) ranks first, and DeepSeek R1 671B is the largest that fits, at 450.9 GB.

512 GB is the most memory a Mac Studio with an M5 Ultra can have, and the biggest size on these pages. It holds the 550B to 671B mixture of experts models on one desk, which otherwise takes a server full of data center GPUs. This page checks every model in the LLM VRAM Calculator against it at Q4_K_M and Q8_0, with 8K and 32K tokens of context.

macOS and your apps share that memory, so the calculator keeps 8 GB for them and checks each model against 504 GB. That's a planning estimate, not a limit Apple sets. Each total counts the weights, the KV cache and a 10% allowance for the runtime.

Models that fit on a 512 GB Mac

Best pick first, by size and release date. Each cell is the total memory the model needs. Tight means under 10% of the memory is left. Pick a number to open it in the calculator.

ModelQ4_K_M8K contextQ4_K_M32K contextQ8_08K contextQ8_032K context
Step-5-Preview 604B (MoE)604B, 27B active, Sep 2026406.1 GBFits408.0 GBFits705.1 GBShort 201.1 GB707.0 GBShort 203.0 GB
Nemotron 3 Ultra 550B A55B (MoE)550B, 55B active, Jun 2026369.6 GBFits370.0 GBFits641.9 GBShort 137.9 GB642.2 GBShort 138.2 GB
Ling-3.1-flash 560B (MoE)560B, 25B active, Oct 2026376.2 GBFits376.5 GBFits653.4 GBShort 149.4 GB653.7 GBShort 149.7 GB
Inkling 2 Small 594B (MoE)594B, 17B active, Nov 2026399.1 GBFits400.0 GBFits693.1 GBShort 189.1 GB694.0 GBShort 190.0 GB
Beam 1.0 501B (MoE)501B, 23B active, Oct 2026336.7 GBFits337.6 GBFits584.6 GBShort 80.6 GB585.6 GBShort 81.6 GB
K2 Horizon 375B A23B (MoE)375B, 23B active, Sep 2026253.9 GBFits260.7 GBFits439.6 GBFits446.3 GBFits
Intern-S2 397B (MoE)397B, 17B active, Sep 2026266.9 GBFits267.7 GBFits463.4 GBTight464.3 GBTight
MiniMax-M3 428B (MoE)428B, 23B active, Jun 2026288.5 GBFits292.2 GBFits500.3 GBTight504.1 GBShort 0.1 GB
GLM-5.3-Flash 320B (MoE)320B, 18B active, Aug 2026215.1 GBFits215.5 GBFits373.5 GBFits373.9 GBFits
MiMo-V2.6-Flash 309B (MoE)309B, 15B active, Sep 2026207.6 GBFits208.2 GBFits360.6 GBFits361.2 GBFits
ERNIE 5.5 Flash 280B (MoE)280B, 14B active, Oct 2026188.0 GBFits188.3 GBFits326.6 GBFits326.9 GBFits
Mistral Medium 3.5 128B128B, Mar 202689.2 GBFits98.9 GBFits152.5 GBFits162.3 GBFits
Arcee Trinity-Large 399B (MoE)399B, 13B active, Apr 2026269.2 GBFits270.8 GBFits466.7 GBTight468.3 GBTight
Llama 3.1 405B405B, Jul 2024276.5 GBFits290.4 GBFits476.9 GBTight490.9 GBTight
MiMo-V2.5 310B (MoE)310B, 15B active, Apr 2026208.3 GBFits208.9 GBFits361.7 GBFits362.4 GBFits

Just too big at Q4_K_M with 8K context

Totals include the weights, the KV cache at FP16 with a batch size of 1, and a 10% allowance for the runtime.

This 512 GB Mac list, in the model finder. Change the memory, quant, context or KV cache type there.
Edit in the full tool

What fits at Q4_K_M

Step-5-Preview 604B (MoE) ranks first at 406.1 GB, with room to spare. The rest of the 550B to 600B class fits too: Nemotron 3 Ultra 550B needs 369.6 GB, Ling-3.1-flash 560B 376.2 GB and Inkling 2 Small 594B 399.1 GB.

The largest model that fits is DeepSeek R1 671B, at 450.9 GB, and DeepSeek V3 671B needs the same. The largest dense model is Llama 3.1 405B, at 276.5 GB.

The order follows the model finder's best pick, which weighs size against release date, so newer models rank above older ones of a similar size.

Long context

Most of the big MoE models here barely grow at 32K. DeepSeek R1 compresses its KV cache, which takes 2.4 GB at 32K, so it needs 452.8 GB there. Llama 3.1 405B keeps a full cache on all 126 of its layers, so its cache is 17.0 GB at 32K. It still fits, at 290.4 GB.

For 64K or 128K, open the model finder from the card below the table and change the context there. Long prompts take a while to read on a Mac, even when the model fits.

Q8_0 on 512 GB

At Q8_0, Llama 3.1 405B fits, with 476.9 GB at 8K and 490.9 GB at 32K, which is tight. MiniMax-M3 428B (MoE) fits at Q8_0 with 8K, at 500.3 GB, also tight.

The 600B and 671B models only fit at Q4_K_M. DeepSeek R1 needs 783.1 GB at Q8_0. In all, 82 models fit at Q8_0 with 8K, against 89 at Q4_K_M.

Just out of reach

GLM-5.3 755B (MoE) misses by 3.5 GB at Q4_K_M, with 507.5 GB, and GLM-5.1 and GLM-5.2 miss too. At Q3_K_M, GLM-5.3 needs 407.9 GB and fits with room.

The 1T models are out of reach even at Q3_K_M. Kimi K2.6 1T (MoE) needs 554.2 GB there. Those take more than one machine, like two Macs with a runner that splits a model across them, or a server with several data center GPUs.

The Mac Studio with 512 GB

Apple sells 512 GB with the M5 Ultra in the Mac Studio. The Ultra has the most memory bandwidth of any Mac. Even so, the 550B to 671B MoE models read 17B to 55B parameters for each token, so they write text more slowly than a small model, and loading one from disk takes a while.

macOS keeps part of the memory away from the GPU by default. On macOS Sonoma and later, this command lets the GPU use 504 GB until the next restart:

bash
sudo sysctl iogpu.wired_limit_mb=516096

Keep an eye on memory pressure in Activity Monitor. If the Mac starts swapping to disk, lower the limit or pick a smaller model.

Other memory sizes

Frequently asked questions

What's the best LLM for a 512 GB Mac?
Step-5-Preview 604B (MoE) ranks first at Q4_K_M with 8K context, at 406.1 GB. DeepSeek R1 671B is the largest that fits, at 450.9 GB. Try a few from the top of the table on your own tasks.
Can a 512 GB Mac run DeepSeek R1 671B?
Yes, at Q4_K_M. DeepSeek R1 671B needs 450.9 GB with 8K and 452.8 GB with 32K. Q8_0 doesn't fit: it needs 783.1 GB.
Can a 512 GB Mac run Kimi K2.6?
No. Kimi K2.6 1T (MoE) needs 689.8 GB at Q4_K_M and 554.2 GB at Q3_K_M. It takes two Macs or a multi GPU server.
Can a 512 GB Mac run Llama 3.1 405B at Q8_0?
Yes. Llama 3.1 405B needs 476.9 GB at Q8_0 with 8K and 490.9 GB with 32K. It's dense, so every token reads all 405B parameters, and replies come slowly.
How much memory does macOS keep on a 512 GB Mac?
The calculator keeps 8 GB for macOS and your apps and gives models 504 GB. That's a planning estimate, not a rule from Apple. macOS also limits the GPU's share by default, and the section above shows how to raise it.

Related tools

LLM VRAM Calculator

Calculate the VRAM needed to run or fine-tune any LLM at any quantization.

Use tool

LLM API Pricing Calculator

Compare API costs across major models (OpenAI, Anthropic, Google) with prompt caching.

Use tool

LLM Token Counter

Count tokens in any prompt for GPT, Claude, Gemini, and Llama with exact OpenAI tokenization.

Use tool

Power Cost Estimator

Estimate annual electricity costs for your PC, Server, or TV.

Use tool