@meta
  v: 1
  route: /gpu/m3-ultra-512
  generated: 2026-06-10T09:07:43.235Z

@intent
  purpose:    Show what AI models fit on the M3 Ultra 512.
  audience:   ai-engineer, self-hoster, homelab-buyer
  capability: inspect_specs, find_fitting_models, compare_to_other_gpus

@state
  slug: m3-ultra-512
  name: M3 Ultra 512
  vram_gb: 512
  bandwidth_gbs: 819
  fp16_tflops: 80
  budget_8k_gb: 413
  category: apple
  summary: "512GB unified memory in a Mac Studio. Runs DeepSeek-V3 at Q4. Yes, really."
  top_fitting_models[5]: deepseek-v3, deepseek-r1, llama-3-1-405b, mixtral-8x22b, gpt-oss-120b

@actions
  - id: compare
    method: GET
    href: /gpu/b200
  - id: compare
    method: GET
    href: /gpu/m2-ultra-192
  - id: compare
    method: GET
    href: /gpu/mi300x
  - id: compare
    method: GET
    href: /gpu/h100-nvl-2x
  - id: view_calculator
    method: GET
    href: /#calculator
  - id: view_math
    method: GET
    href: /the-math

@context
  > 512GB unified memory in a Mac Studio. Runs DeepSeek-V3 at Q4. Yes, really.

@nav
  self:      /gpu/m3-ultra-512
  parents:   [/]
  peers:     [/gpu/b200, /gpu/m2-ultra-192, /gpu/mi300x, /gpu/h100-nvl-2x]
  drilldown: /the-math