What LLMs Can the Apple M3 Ultra 512GB Run?

The Apple M3 Ultra 512GB has 512 GB of unified memory (about 384 GB usable by the GPU) and 819 GB/s of bandwidth. It runs 56 of 58 models fully at Q4_K_M with 8K context, up to Llama 3.1 405B, and generates about 95 tokens/s on an 8B model.

Apple · Apple Silicon · Released 2025-03 · Updated October 14, 2026

57 of 58 models fit

Apple M3 Ultra 512GB (384 GB GPU-usable) at Q4_K_M with 8K context. Largest that fits fully: DeepSeek R1 671B.

ModelNeedsStatusSpeedBest quant that fits
Llama 3.1 405B234 GBFits~2.1 tok/sQ6_K
Qwen 3.5 397B-A17B226 GBFits~39 tok/sQ6_K
DeepSeek V4 Flash161 GBFits~52 tok/sFP4 + FP8 (official)
Qwen 3 235B-A22B135 GBFits~28 tok/sQ8_0
MiniMax-M2133 GBFits~51 tok/sFP8 (official)
Qwen 3.5 122B-A10B70.2 GBFits~60 tok/sFP16 / BF16
Mistral Small 4 119B68.2 GBFits~89 tok/sFP8 (official)
gpt-oss-120b67.2 GBFits~103 tok/sMXFP4 (official)
Llama 4 Scout64.0 GBFits~39 tok/sFP16 / BF16
GLM-4.5-Air62.1 GBFits~51 tok/sFP16 / BF16
Qwen3-Next 80B-A3B46.3 GBFits~117 tok/sFP16 / BF16
Qwen 2.5 72B44.5 GBFits~12 tok/sFP16 / BF16
Llama 3.3 70B43.2 GBFits~12 tok/sFP16 / BF16
DeepSeek R1 Distill Llama 70B43.2 GBFits~12 tok/sFP16 / BF16
Mixtral 8x7B28.0 GBFits~53 tok/sFP16 / BF16
Qwen 2.5 32B21.4 GBFits~25 tok/sFP16 / BF16
Qwen 2.5 Coder 32B21.4 GBFits~25 tok/sFP16 / BF16
DeepSeek R1 Distill Qwen 32B21.4 GBFits~25 tok/sFP16 / BF16
Qwen 3 32B21.4 GBFits~25 tok/sFP16 / BF16
Qwen 3.5 35B-A3B21.0 GBFits~130 tok/sFP16 / BF16
Qwen 3.6 35B-A3B21.0 GBFits~130 tok/sFP16 / BF16
Gemma 4 31B19.8 GBFits~26 tok/sFP16 / BF16
Qwen 3 30B-A3B18.8 GBFits~107 tok/sFP16 / BF16
Qwen3-Coder 30B-A3B18.8 GBFits~107 tok/sFP16 / BF16
Gemma 3 27B17.6 GBFits~30 tok/sFP16 / BF16
Qwen 3.5 27B16.9 GBFits~31 tok/sFP16 / BF16
Qwen 3.6 27B16.9 GBFits~31 tok/sFP16 / BF16
Qwen 3.8 27B16.9 GBFits~31 tok/sFP16 / BF16
Mistral Small 3.1 24B15.6 GBFits~34 tok/sFP16 / BF16
Gemma 4 26B-A4B15.6 GBFits~127 tok/sFP16 / BF16
gpt-oss-20b12.9 GBFits~149 tok/sMXFP4 (official)
Qwen 2.5 14B10.7 GBFits~53 tok/sFP16 / BF16
DeepSeek R1 Distill Qwen 14B10.7 GBFits~53 tok/sFP16 / BF16
Phi-4 14B10.6 GBFits~53 tok/sFP16 / BF16
Qwen 3 14B10.5 GBFits~54 tok/sFP16 / BF16
Ministral 3 14B9.7 GBFits~59 tok/sFP16 / BF16
Mistral Nemo 12B9.0 GBFits~64 tok/sFP16 / BF16
Gemma 3 12B8.8 GBFits~64 tok/sFP16 / BF16
Ministral 3 8B6.7 GBFits~90 tok/sFP16 / BF16
Qwen 3 8B6.6 GBFits~92 tok/sFP16 / BF16
Qwen 3.5 9B6.4 GBFits~92 tok/sFP16 / BF16
Llama 3.1 8B6.3 GBFits~95 tok/sFP16 / BF16
DeepSeek R1 Distill Llama 8B6.3 GBFits~95 tok/sFP16 / BF16
Mistral 7B v0.35.7 GBFits~104 tok/sFP16 / BF16
Gemma 4 E4B5.7 GBFits~103 tok/sFP16 / BF16
Qwen 2.5 7B5.6 GBFits~106 tok/sFP16 / BF16
DeepSeek R1 Distill Qwen 7B5.6 GBFits~106 tok/sFP16 / BF16
Qwen 3 4B4.2 GBFits~165 tok/sFP16 / BF16
Phi-4 Mini4.1 GBFits~175 tok/sFP16 / BF16
Gemma 4 E2B4.0 GBFits~161 tok/sFP16 / BF16
Gemma 3 4B3.7 GBFits~178 tok/sFP16 / BF16
Qwen 3.5 4B3.7 GBFits~183 tok/sFP16 / BF16
Ministral 3 3B3.5 GBFits~201 tok/sFP16 / BF16
Llama 3.2 3B3.5 GBFits~208 tok/sFP16 / BF16
Llama 3.2 1B1.7 GBFits~548 tok/sFP16 / BF16
Gemma 3 1B1.6 GBFits~689 tok/sFP16 / BF16
DeepSeek R1 671B380 GBTight fit~21 tok/sIQ4_XS
Kimi K2.6566 GBToo small–Q2_K

Apple M3 Ultra 512GB for local AI

The Mac Studio configuration that runs the largest open models in one box, including DeepSeek R1 at Q4_K_M once you raise the default GPU memory limit (about 75% of RAM). Speed is limited by its 819 GB/s bandwidth.

Setup tips: with about 384 GB GPU-usable by default (more with a raised wired limit) it can hold Qwen 3.5 397B, MiniMax-M2 or DeepSeek V4 Flash. Long prompts process slowly compared with NVIDIA GPUs, so keep system prompts short and reuse prompt caches where your app supports it.

Our picks for the Apple M3 Ultra 512GB

The model we would run first for each job, chosen from our recommendations by what fits fully with an 8K context. "Best quant" is the highest quality that fits; "max context" is at Q4_K_M or the native format.

TaskModelBest quantMax contextSpeed
Best all-rounderQwen 3.5 397B-A17BQ6_K256K~39 tok/s
CodingMiniMax-M2FP8 (official)192K~37 tok/s
Fastest repliesQwen 3.5 122B-A10BFP16 / BF16256K~60 tok/s
Long documents (64K+)Qwen 3.5 122B-A10BFP16 / BF16256K~60 tok/s
Images in, text outGemma 4 31BFP16 / BF16256K~26 tok/s
Image generationFLUX.2 [dev]FP8––
SpecValue
Memory512 GB unified (384 GB GPU-usable by default)
Memory bandwidth819 GB/s
ArchitectureApple Silicon
Released2025-03
Llama 3.1 8B Q4_K_M speed~95 tok/s
Largest model that fitsLlama 3.1 405B

Image and video models on the Apple M3 Ultra 512GB

Peak VRAM with model swapping (ComfyUI default on smaller cards).

ModelFP8FP8 fitGGUF Q4Q4 fit
Stable Diffusion 1.51.8 GBFits1.7 GBFits
SDXL 1.04.1 GBFits4.1 GBFits
Stable Diffusion 3.5 Medium6.4 GBFits6.4 GBFits
Stable Diffusion 3.5 Large8.6 GBFits6.4 GBFits
FLUX.1 [dev]12.1 GBFits7.3 GBFits
FLUX.1 [schnell]12.1 GBFits7.3 GBFits
Qwen-Image20.4 GBFits12.3 GBFits
Wan 2.2 A14B (T2V / I2V)18.8 GBFits13.2 GBFits
Wan 2.2 TI2V 5B8.0 GBFits6.5 GBFits
HunyuanVideo18.3 GBFits13.1 GBFits
FLUX.1 Kontext [dev]12.3 GBFits7.5 GBFits
FLUX.2 [dev]31.0 GBFits23.6 GBFits
Z-Image Turbo6.8 GBFits4.9 GBFits
Qwen-Image-Edit 250920.5 GBFits12.4 GBFits
LTX-219.2 GBFits12.6 GBFits

Comparisons

Apple M3 Ultra 512GB for LLMs: common questions

Can the Apple M3 Ultra 512GB run Llama 3.3 70B?

Yes. Llama 3.3 70B needs about 43.2 GB at Q4_K_M with 8K context, and the best quality that fits is FP16 / BF16.

Can the Apple M3 Ultra 512GB run Qwen 3 32B?

Yes. Qwen 3 32B needs about 21.4 GB at Q4_K_M with 8K context, and the best quality that fits is FP16 / BF16.

Can the Apple M3 Ultra 512GB run Qwen 3 14B?

Yes. Qwen 3 14B needs about 10.5 GB at Q4_K_M with 8K context, and the best quality that fits is FP16 / BF16.

How fast is the Apple M3 Ultra 512GB for LLMs?

About 95 tokens per second for an 8B model at Q4_K_M, based on its 819 GB/s memory bandwidth. Larger models scale down roughly in proportion to their size.

How much of the 512 GB can the GPU use?

About 384 GB by default. You can raise it with sudo sysctl iogpu.wired_limit_mb=<MB> (resets on reboot), leaving at least 8 GB for macOS.

Related