Skip to content
Run LLM Locally

Mac · Apple Silicon

M3 Ultra (80-core GPU) with 512 GB: which local LLMs run?

With 512 GB of unified memory, the M3 Ultra (80-core GPU) gives the GPU about 384 GB by default (819 GB/s): enough for models up to about DeepSeek R1 (671B).

  • Unified memory512 GB
  • GPU usable~384 GB
  • Bandwidth819 GB/s
  • GPU cores80
ShareFacebookXWhatsAppLinkedInEmail

Unified memory is not VRAM: macOS reserves part of it. The limit can be raised with sudo sysctl iogpu.wired_limit_mb=…, at the cost of memory for apps. There is no offload to system RAM — it is the same memory.

CPU cores
32
GPU cores
80
Bandwidth
819 GB/s
Memory configurations
96 GB · 256 GB · 512 GB
Sold in
Mac Studio
Released
March 2025

Official specData updated 24 days agoSource: Apple tech specs

What the M3 Ultra (80-core GPU) with 512 GB runs

M3 Ultra (80-core GPU) · 512 GB · 8K
Qwen3 235B-A22BQ8_0 · 8K~247 GB · 14–21 tok/s estimateExcellent
gpt-oss-120bMXFP4 · 8K~63 GB · 70–92 tok/s estimateExcellent
Llama 4 Scout 17B-16E InstructQ8_0 · 8K~113 GB · 20–30 tok/s estimateExcellent
Mixtral 8x7B InstructQ8_0 · 8K~50 GB · 27–40 tok/s estimateExcellent
Qwen3 32BQ8_0 · 8K~36.8 GB · 14–21 tok/s estimateExcellent
DeepSeek R1 Distill Qwen 32BQ8_0 · 8K~36.8 GB · 14–21 tok/s estimateExcellent
Qwen2.5 32B InstructQ8_0 · 8K~36.8 GB · 14–21 tok/s estimateExcellent
Qwen2.5 Coder 32B InstructQ8_0 · 8K~36.8 GB · 14–21 tok/s estimateExcellent
QwQ 32BQ8_0 · 8K~36.8 GB · 14–21 tok/s estimateExcellent
Qwen3 30B-A3BQ8_0 · 8K~33.2 GB · 51–89 tok/s estimateExcellent
Qwen3 Coder 30B-A3B InstructQ8_0 · 8K~33.2 GB · 51–89 tok/s estimateExcellent
Gemma 3 27B ITQ8_0 · 8K~30 GB · 17–26 tok/s estimateExcellent
Gemma 2 27B ITQ8_0 · 8K~31.4 GB · 17–25 tok/s estimateExcellent
Mistral Small 3.2 24B InstructQ8_0 · 8K~26.4 GB · 20–29 tok/s estimateExcellent
gpt-oss-20bMXFP4 · 8K~12.8 GB · 105–116 tok/s measuredExcellent
Qwen3 14BQ8_0 · 8K~17.3 GB · 29–46 tok/s estimateExcellent
DeepSeek R1 Distill Qwen 14BQ8_0 · 8K~17.6 GB · 29–46 tok/s estimateExcellent
Qwen2.5 14B InstructQ8_0 · 8K~17.6 GB · 29–46 tok/s estimateExcellent
Phi-4 14BQ8_0 · 8K~17.4 GB · 29–46 tok/s estimateExcellent
Mistral Nemo 12B InstructQ8_0 · 8K~14.7 GB · 34–55 tok/s estimateExcellent
Gemma 3 12B ITQ8_0 · 8K~13.9 GB · 35–58 tok/s estimateExcellent
Gemma 2 9B ITQ8_0 · 8K~12.5 GB · 41–69 tok/s estimateExcellent
Qwen3 8BQ8_0 · 8K~10.3 GB · 48–81 tok/s estimateExcellent
Llama 3.1 8B InstructQ8_0 · 8K~10 GB · 50–83 tok/s estimateExcellent
DeepSeek R1 Distill Llama 8BQ8_0 · 8K~10 GB · 50–83 tok/s estimateExcellent
Qwen2.5 7B InstructQ8_0 · 8K~9 GB · 55–91 tok/s estimateExcellent
Mistral 7B Instruct v0.3Q8_0 · 8K~9 GB · 55–92 tok/s estimateExcellent
Gemma 3 4B ITQ8_0 · 8K~5.2 GB · 89–176 tok/s estimateExcellent
Qwen3 4BQ8_0 · 8K~6 GB · 80–155 tok/s estimateExcellent
Phi-4 Mini Instruct 3.8BQ8_0 · 8K~5.8 GB · 86–164 tok/s estimateExcellent
Llama 3.2 3B InstructQ8_0 · 8K~4.9 GB · 101–194 tok/s estimateExcellent
Qwen3 1.7BQ8_0 · 8K~3.5 GB · 142–328 tok/s estimateExcellent
Llama 3.2 1B InstructQ8_0 · 8K~2.3 GB · 234–520 tok/s estimateExcellent
Qwen3 0.6BQ8_0 · 8K~2.3 GB · 207–682 tok/s estimateExcellent
DeepSeek R1 (671B)IQ4_XS · 8K~353 GB · 18–27 tok/s estimateCompatible
Qwen2.5 72B InstructQ8_0 · 8K~79.1 GB · 6.5–9.5 tok/s estimateCompatible
Llama 3.3 70B InstructQ8_0 · 8K~76.6 GB · 7–10 tok/s estimateCompatible
DeepSeek R1 Distill Llama 70BQ8_0 · 8K~76.6 GB · 7–10 tok/s estimateCompatible

For each model, the highest-precision quantization that runs entirely in fast memory; otherwise the one with the best result. Speeds are ranges; "estimate" unless a benchmark exists for this exact hardware.

EstimateData updated 24 days agoSource: Run LLM Locally engine

Where to buy

Prices and links follow the market you choose.

Mac Studio M3 Ultra

M3 Ultra (80-core GPU) · 819 GB/s

Price not available

Availability unknown

View on amazon.com

Affiliate link

No current offer imported for this market: the link opens a marketplace search.