[
  {
    "id": "mac-mini-m6-16",
    "family": "Mac mini",
    "chip": "M6",
    "chip_variant": "12-core CPU / 12-core GPU",
    "unified_memory_gb": 16,
    "memory_bandwidth_gbs": 153,
    "usable_memory_gb": 10.5,
    "price_usd": 899,
    "idle_watts": null,
    "load_watts": 65,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has not published power figures for the 2026 machines yet (support articles 103253 / 102027 stop at the 2024/2025 models). Stand-in: Apple's published maximum for the previous chip (M4 Mac mini, 65 W), which is a CPU+GPU stress figure, so it likely overstates inference draw.",
    "generation": "current",
    "status": "Pre-order; ships 2026-09-22.",
    "notes": "macOS lets the GPU wire roughly two-thirds of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://www.apple.com/mac-mini/specs/",
      "https://www.apple.com/newsroom/2026/08/apple-unveils-a-more-powerful-mac-mini-featuring-the-all-new-m6-and-m5-pro/",
      "https://daringfireball.net/2026/08/configurations_and_pricing_for_new_mac_minis_and_mac_studios"
    ]
  },
  {
    "id": "mac-mini-m6-24",
    "family": "Mac mini",
    "chip": "M6",
    "chip_variant": "12-core CPU / 12-core GPU",
    "unified_memory_gb": 24,
    "memory_bandwidth_gbs": 170,
    "usable_memory_gb": 16.0,
    "price_usd": 1099,
    "idle_watts": null,
    "load_watts": 65,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has not published power figures for the 2026 machines yet (support articles 103253 / 102027 stop at the 2024/2025 models). Stand-in: Apple's published maximum for the previous chip (M4 Mac mini, 65 W), which is a CPU+GPU stress figure, so it likely overstates inference draw.",
    "generation": "current",
    "status": "Pre-order; ships 2026-09-22.",
    "notes": "macOS lets the GPU wire roughly two-thirds of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://www.apple.com/mac-mini/specs/",
      "https://www.apple.com/newsroom/2026/08/apple-unveils-a-more-powerful-mac-mini-featuring-the-all-new-m6-and-m5-pro/",
      "https://daringfireball.net/2026/08/configurations_and_pricing_for_new_mac_minis_and_mac_studios"
    ]
  },
  {
    "id": "mac-mini-m6-32",
    "family": "Mac mini",
    "chip": "M6",
    "chip_variant": "12-core CPU / 12-core GPU",
    "unified_memory_gb": 32,
    "memory_bandwidth_gbs": 170,
    "usable_memory_gb": 21.0,
    "price_usd": 1299,
    "idle_watts": null,
    "load_watts": 65,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has not published power figures for the 2026 machines yet (support articles 103253 / 102027 stop at the 2024/2025 models). Stand-in: Apple's published maximum for the previous chip (M4 Mac mini, 65 W), which is a CPU+GPU stress figure, so it likely overstates inference draw.",
    "generation": "current",
    "status": "Pre-order; ships 2026-09-22.",
    "notes": "macOS lets the GPU wire roughly two-thirds of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://www.apple.com/mac-mini/specs/",
      "https://www.apple.com/newsroom/2026/08/apple-unveils-a-more-powerful-mac-mini-featuring-the-all-new-m6-and-m5-pro/",
      "https://daringfireball.net/2026/08/configurations_and_pricing_for_new_mac_minis_and_mac_studios"
    ]
  },
  {
    "id": "mac-mini-m5-pro-24",
    "family": "Mac mini",
    "chip": "M5 Pro",
    "chip_variant": "15-core CPU / 16-core GPU",
    "unified_memory_gb": 24,
    "memory_bandwidth_gbs": 307,
    "usable_memory_gb": 16.0,
    "price_usd": 1699,
    "idle_watts": null,
    "load_watts": 140,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has not published power figures for the 2026 machines yet (support articles 103253 / 102027 stop at the 2024/2025 models). Stand-in: Apple's published maximum for the previous chip (M4 Pro Mac mini, 140 W), which is a CPU+GPU stress figure, so it likely overstates inference draw.",
    "generation": "current",
    "status": "Pre-order; ships 2026-09-22.",
    "notes": "macOS lets the GPU wire roughly two-thirds of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://www.apple.com/mac-mini/specs/",
      "https://www.apple.com/newsroom/2026/08/apple-unveils-a-more-powerful-mac-mini-featuring-the-all-new-m6-and-m5-pro/",
      "https://daringfireball.net/2026/08/configurations_and_pricing_for_new_mac_minis_and_mac_studios"
    ]
  },
  {
    "id": "mac-mini-m5-pro-48",
    "family": "Mac mini",
    "chip": "M5 Pro",
    "chip_variant": "15-core CPU / 16-core GPU",
    "unified_memory_gb": 48,
    "memory_bandwidth_gbs": 307,
    "usable_memory_gb": 36.0,
    "price_usd": 2299,
    "idle_watts": null,
    "load_watts": 140,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has not published power figures for the 2026 machines yet (support articles 103253 / 102027 stop at the 2024/2025 models). Stand-in: Apple's published maximum for the previous chip (M4 Pro Mac mini, 140 W), which is a CPU+GPU stress figure, so it likely overstates inference draw.",
    "generation": "current",
    "status": "Pre-order; ships 2026-09-22.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://www.apple.com/mac-mini/specs/",
      "https://www.apple.com/newsroom/2026/08/apple-unveils-a-more-powerful-mac-mini-featuring-the-all-new-m6-and-m5-pro/",
      "https://daringfireball.net/2026/08/configurations_and_pricing_for_new_mac_minis_and_mac_studios"
    ]
  },
  {
    "id": "mac-mini-m5-pro-64",
    "family": "Mac mini",
    "chip": "M5 Pro",
    "chip_variant": "15-core CPU / 16-core GPU",
    "unified_memory_gb": 64,
    "memory_bandwidth_gbs": 307,
    "usable_memory_gb": 48.0,
    "price_usd": 2699,
    "idle_watts": null,
    "load_watts": 140,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has not published power figures for the 2026 machines yet (support articles 103253 / 102027 stop at the 2024/2025 models). Stand-in: Apple's published maximum for the previous chip (M4 Pro Mac mini, 140 W), which is a CPU+GPU stress figure, so it likely overstates inference draw.",
    "generation": "current",
    "status": "Pre-order; ships 2026-09-22.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://www.apple.com/mac-mini/specs/",
      "https://www.apple.com/newsroom/2026/08/apple-unveils-a-more-powerful-mac-mini-featuring-the-all-new-m6-and-m5-pro/",
      "https://daringfireball.net/2026/08/configurations_and_pricing_for_new_mac_minis_and_mac_studios"
    ]
  },
  {
    "id": "mac-studio-m5-max-36",
    "family": "Mac Studio",
    "chip": "M5 Max",
    "chip_variant": "18-core CPU / 32-core GPU",
    "unified_memory_gb": 36,
    "memory_bandwidth_gbs": 460,
    "usable_memory_gb": 27.0,
    "price_usd": 2499,
    "idle_watts": null,
    "load_watts": 145,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has not published power figures for the 2026 machines yet (support articles 103253 / 102027 stop at the 2024/2025 models). Stand-in: Apple's published maximum for the previous chip (M4 Max Mac Studio, 145 W), which is a CPU+GPU stress figure, so it likely overstates inference draw.",
    "generation": "current",
    "status": "Pre-order; ships 2026-09-22.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://www.apple.com/mac-studio/specs/",
      "https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/",
      "https://daringfireball.net/2026/08/configurations_and_pricing_for_new_mac_minis_and_mac_studios"
    ]
  },
  {
    "id": "mac-studio-m5-max-48",
    "family": "Mac Studio",
    "chip": "M5 Max",
    "chip_variant": "18-core CPU / 40-core GPU",
    "unified_memory_gb": 48,
    "memory_bandwidth_gbs": 614,
    "usable_memory_gb": 36.0,
    "price_usd": 3099,
    "idle_watts": null,
    "load_watts": 145,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has not published power figures for the 2026 machines yet (support articles 103253 / 102027 stop at the 2024/2025 models). Stand-in: Apple's published maximum for the previous chip (M4 Max Mac Studio, 145 W), which is a CPU+GPU stress figure, so it likely overstates inference draw.",
    "generation": "current",
    "status": "Pre-order; ships 2026-09-22.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://www.apple.com/mac-studio/specs/",
      "https://daringfireball.net/2026/08/configurations_and_pricing_for_new_mac_minis_and_mac_studios"
    ]
  },
  {
    "id": "mac-studio-m5-max-64",
    "family": "Mac Studio",
    "chip": "M5 Max",
    "chip_variant": "18-core CPU / 40-core GPU",
    "unified_memory_gb": 64,
    "memory_bandwidth_gbs": 614,
    "usable_memory_gb": 48.0,
    "price_usd": 3499,
    "idle_watts": null,
    "load_watts": 145,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has not published power figures for the 2026 machines yet (support articles 103253 / 102027 stop at the 2024/2025 models). Stand-in: Apple's published maximum for the previous chip (M4 Max Mac Studio, 145 W), which is a CPU+GPU stress figure, so it likely overstates inference draw.",
    "generation": "current",
    "status": "Pre-order; ships 2026-09-22.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://www.apple.com/mac-studio/specs/",
      "https://daringfireball.net/2026/08/configurations_and_pricing_for_new_mac_minis_and_mac_studios"
    ]
  },
  {
    "id": "mac-studio-m5-max-128",
    "family": "Mac Studio",
    "chip": "M5 Max",
    "chip_variant": "18-core CPU / 40-core GPU",
    "unified_memory_gb": 128,
    "memory_bandwidth_gbs": 614,
    "usable_memory_gb": 96.0,
    "price_usd": 5099,
    "idle_watts": null,
    "load_watts": 145,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has not published power figures for the 2026 machines yet (support articles 103253 / 102027 stop at the 2024/2025 models). Stand-in: Apple's published maximum for the previous chip (M4 Max Mac Studio, 145 W), which is a CPU+GPU stress figure, so it likely overstates inference draw.",
    "generation": "current",
    "status": "Pre-order; ships 2026-09-22.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://www.apple.com/mac-studio/specs/",
      "https://daringfireball.net/2026/08/configurations_and_pricing_for_new_mac_minis_and_mac_studios"
    ]
  },
  {
    "id": "mac-studio-m5-ultra-96",
    "family": "Mac Studio",
    "chip": "M5 Ultra",
    "chip_variant": "30-core CPU / 64-core GPU",
    "unified_memory_gb": 96,
    "memory_bandwidth_gbs": 1200,
    "usable_memory_gb": 72.0,
    "price_usd": 5499,
    "idle_watts": null,
    "load_watts": 270,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has not published power figures for the 2026 machines yet (support articles 103253 / 102027 stop at the 2024/2025 models). Stand-in: Apple's published maximum for the previous chip (M3 Ultra Mac Studio, 270 W), which is a CPU+GPU stress figure, so it likely overstates inference draw.",
    "generation": "current",
    "status": "Pre-order; ships 2026-09-22.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://www.apple.com/mac-studio/specs/",
      "https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/",
      "https://daringfireball.net/2026/08/configurations_and_pricing_for_new_mac_minis_and_mac_studios"
    ]
  },
  {
    "id": "mac-studio-m5-ultra-256",
    "family": "Mac Studio",
    "chip": "M5 Ultra",
    "chip_variant": "36-core CPU / 80-core GPU",
    "unified_memory_gb": 256,
    "memory_bandwidth_gbs": 1200,
    "usable_memory_gb": 192.0,
    "price_usd": 10799,
    "idle_watts": null,
    "load_watts": 270,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has not published power figures for the 2026 machines yet (support articles 103253 / 102027 stop at the 2024/2025 models). Stand-in: Apple's published maximum for the previous chip (M3 Ultra Mac Studio, 270 W), which is a CPU+GPU stress figure, so it likely overstates inference draw.",
    "generation": "current",
    "status": "Pre-order; ships 2026-09-22.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://www.apple.com/mac-studio/specs/",
      "https://daringfireball.net/2026/08/configurations_and_pricing_for_new_mac_minis_and_mac_studios",
      "https://appleinsider.com/articles/26/08/25/you-can-spend-18299-on-a-mac-studio-today-or-more-in-october"
    ]
  },
  {
    "id": "mac-studio-m5-ultra-512",
    "family": "Mac Studio",
    "chip": "M5 Ultra",
    "chip_variant": "36-core CPU / 80-core GPU",
    "unified_memory_gb": 512,
    "memory_bandwidth_gbs": 1200,
    "usable_memory_gb": 384.0,
    "price_usd": null,
    "idle_watts": null,
    "load_watts": 270,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has not published power figures for the 2026 machines yet (support articles 103253 / 102027 stop at the 2024/2025 models). Stand-in: Apple's published maximum for the previous chip (M3 Ultra Mac Studio, 270 W), which is a CPU+GPU stress figure, so it likely overstates inference draw.",
    "generation": "current",
    "status": "Announced for late October 2026; Apple has not published a price.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://www.apple.com/mac-studio/specs/",
      "https://www.apple.com/mac-studio/"
    ],
    "TODO": "price_usd: Apple has not announced the 512GB price."
  },
  {
    "id": "mac-mini-m4-16",
    "family": "Mac mini",
    "chip": "M4",
    "chip_variant": "10-core CPU / 10-core GPU",
    "unified_memory_gb": 16,
    "memory_bandwidth_gbs": 120,
    "usable_memory_gb": 10.5,
    "price_usd": 599,
    "idle_watts": 4,
    "load_watts": 65,
    "load_watts_status": "published",
    "load_watts_note": "Apple's published 'max' figure for this chip (support article 103253), a CPU+GPU stress number; inference typically draws less.",
    "generation": "previous",
    "status": "Discontinued 2026-08-25. Launch price; Apple no longer sells it new, so treat as a refurbished/used reference.",
    "notes": "macOS lets the GPU wire roughly two-thirds of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://support.apple.com/en-us/121555",
      "https://support.apple.com/en-us/103253",
      "https://lowendmac.com/2024/mac-mini-2024/"
    ]
  },
  {
    "id": "mac-mini-m4-24",
    "family": "Mac mini",
    "chip": "M4",
    "chip_variant": "10-core CPU / 10-core GPU",
    "unified_memory_gb": 24,
    "memory_bandwidth_gbs": 120,
    "usable_memory_gb": 16.0,
    "price_usd": 799,
    "idle_watts": 4,
    "load_watts": 65,
    "load_watts_status": "published",
    "load_watts_note": "Apple's published 'max' figure for this chip (support article 103253), a CPU+GPU stress number; inference typically draws less.",
    "generation": "previous",
    "status": "Discontinued 2026-08-25. Launch price; Apple no longer sells it new, so treat as a refurbished/used reference.",
    "notes": "macOS lets the GPU wire roughly two-thirds of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://support.apple.com/en-us/121555",
      "https://support.apple.com/en-us/103253",
      "https://lowendmac.com/2024/mac-mini-2024/"
    ]
  },
  {
    "id": "mac-mini-m4-32",
    "family": "Mac mini",
    "chip": "M4",
    "chip_variant": "10-core CPU / 10-core GPU",
    "unified_memory_gb": 32,
    "memory_bandwidth_gbs": 120,
    "usable_memory_gb": 21.0,
    "price_usd": 999,
    "idle_watts": 4,
    "load_watts": 65,
    "load_watts_status": "published",
    "load_watts_note": "Apple's published 'max' figure for this chip (support article 103253), a CPU+GPU stress number; inference typically draws less.",
    "generation": "previous",
    "status": "Discontinued 2026-08-25. Launch price; Apple no longer sells it new, so treat as a refurbished/used reference.",
    "notes": "macOS lets the GPU wire roughly two-thirds of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://support.apple.com/en-us/121555",
      "https://support.apple.com/en-us/103253",
      "https://lowendmac.com/2024/mac-mini-2024/"
    ]
  },
  {
    "id": "mac-mini-m4-pro-24",
    "family": "Mac mini",
    "chip": "M4 Pro",
    "chip_variant": "12-core CPU / 16-core GPU",
    "unified_memory_gb": 24,
    "memory_bandwidth_gbs": 273,
    "usable_memory_gb": 16.0,
    "price_usd": 1399,
    "idle_watts": 5,
    "load_watts": 140,
    "load_watts_status": "published",
    "load_watts_note": "Apple's published 'max' figure for this chip (support article 103253), a CPU+GPU stress number; inference typically draws less.",
    "generation": "previous",
    "status": "Discontinued 2026-08-25. Launch price; Apple no longer sells it new, so treat as a refurbished/used reference.",
    "notes": "macOS lets the GPU wire roughly two-thirds of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://support.apple.com/en-us/121555",
      "https://support.apple.com/en-us/103253",
      "https://lowendmac.com/2024/mac-mini-2024/"
    ]
  },
  {
    "id": "mac-mini-m4-pro-48",
    "family": "Mac mini",
    "chip": "M4 Pro",
    "chip_variant": "12-core CPU / 16-core GPU",
    "unified_memory_gb": 48,
    "memory_bandwidth_gbs": 273,
    "usable_memory_gb": 36.0,
    "price_usd": 1799,
    "idle_watts": 5,
    "load_watts": 140,
    "load_watts_status": "published",
    "load_watts_note": "Apple's published 'max' figure for this chip (support article 103253), a CPU+GPU stress number; inference typically draws less.",
    "generation": "previous",
    "status": "Discontinued 2026-08-25. Launch price; Apple no longer sells it new, so treat as a refurbished/used reference.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://support.apple.com/en-us/121555",
      "https://support.apple.com/en-us/103253",
      "https://lowendmac.com/2024/mac-mini-2024/"
    ]
  },
  {
    "id": "mac-mini-m4-pro-64",
    "family": "Mac mini",
    "chip": "M4 Pro",
    "chip_variant": "12-core CPU / 16-core GPU",
    "unified_memory_gb": 64,
    "memory_bandwidth_gbs": 273,
    "usable_memory_gb": 48.0,
    "price_usd": 1999,
    "idle_watts": 5,
    "load_watts": 140,
    "load_watts_status": "published",
    "load_watts_note": "Apple's published 'max' figure for this chip (support article 103253), a CPU+GPU stress number; inference typically draws less.",
    "generation": "previous",
    "status": "Discontinued 2026-08-25. Launch price; Apple no longer sells it new, so treat as a refurbished/used reference.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://support.apple.com/en-us/121555",
      "https://support.apple.com/en-us/103253",
      "https://lowendmac.com/2024/mac-mini-2024/"
    ]
  },
  {
    "id": "mac-studio-m4-max-36",
    "family": "Mac Studio",
    "chip": "M4 Max",
    "chip_variant": "14-core CPU / 32-core GPU",
    "unified_memory_gb": 36,
    "memory_bandwidth_gbs": 410,
    "usable_memory_gb": 27.0,
    "price_usd": 1999,
    "idle_watts": 6,
    "load_watts": 145,
    "load_watts_status": "published",
    "load_watts_note": "Apple's published 'max' figure for this chip (support article 102027), a CPU+GPU stress number; inference typically draws less. Apple measured the 14-core/32-GPU config.",
    "generation": "previous",
    "status": "Discontinued 2026-08-25. Launch price; Apple no longer sells it new, so treat as a refurbished/used reference.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://support.apple.com/en-us/122211",
      "https://support.apple.com/en-us/102027",
      "https://lowendmac.com/2025/mac-studio-early-2025/"
    ]
  },
  {
    "id": "mac-studio-m4-max-48",
    "family": "Mac Studio",
    "chip": "M4 Max",
    "chip_variant": "16-core CPU / 40-core GPU",
    "unified_memory_gb": 48,
    "memory_bandwidth_gbs": 546,
    "usable_memory_gb": 36.0,
    "price_usd": 2499,
    "idle_watts": 6,
    "load_watts": 145,
    "load_watts_status": "published",
    "load_watts_note": "Apple's published 'max' figure for this chip (support article 102027), a CPU+GPU stress number; inference typically draws less. Apple measured the 14-core/32-GPU config only.",
    "generation": "previous",
    "status": "Discontinued 2026-08-25. Launch price; Apple no longer sells it new, so treat as a refurbished/used reference.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://support.apple.com/en-us/122211",
      "https://support.apple.com/en-us/102027",
      "https://lowendmac.com/2025/mac-studio-early-2025/"
    ]
  },
  {
    "id": "mac-studio-m4-max-64",
    "family": "Mac Studio",
    "chip": "M4 Max",
    "chip_variant": "16-core CPU / 40-core GPU",
    "unified_memory_gb": 64,
    "memory_bandwidth_gbs": 546,
    "usable_memory_gb": 48.0,
    "price_usd": 2699,
    "idle_watts": 6,
    "load_watts": 145,
    "load_watts_status": "published",
    "load_watts_note": "Apple's published 'max' figure for this chip (support article 102027), a CPU+GPU stress number; inference typically draws less. Apple measured the 14-core/32-GPU config only.",
    "generation": "previous",
    "status": "Discontinued 2026-08-25. Launch price; Apple no longer sells it new, so treat as a refurbished/used reference.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://support.apple.com/en-us/122211",
      "https://support.apple.com/en-us/102027",
      "https://lowendmac.com/2025/mac-studio-early-2025/"
    ]
  },
  {
    "id": "mac-studio-m4-max-128",
    "family": "Mac Studio",
    "chip": "M4 Max",
    "chip_variant": "16-core CPU / 40-core GPU",
    "unified_memory_gb": 128,
    "memory_bandwidth_gbs": 546,
    "usable_memory_gb": 96.0,
    "price_usd": 3499,
    "idle_watts": 6,
    "load_watts": 145,
    "load_watts_status": "published",
    "load_watts_note": "Apple's published 'max' figure for this chip (support article 102027), a CPU+GPU stress number; inference typically draws less. Apple measured the 14-core/32-GPU config only.",
    "generation": "previous",
    "status": "Discontinued 2026-08-25. Launch price; Apple no longer sells it new, so treat as a refurbished/used reference.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://support.apple.com/en-us/122211",
      "https://support.apple.com/en-us/102027",
      "https://lowendmac.com/2025/mac-studio-early-2025/"
    ]
  },
  {
    "id": "mac-studio-m3-ultra-96",
    "family": "Mac Studio",
    "chip": "M3 Ultra",
    "chip_variant": "28-core CPU / 60-core GPU",
    "unified_memory_gb": 96,
    "memory_bandwidth_gbs": 819,
    "usable_memory_gb": 72.0,
    "price_usd": 3999,
    "idle_watts": 9,
    "load_watts": 270,
    "load_watts_status": "published",
    "load_watts_note": "Apple's published 'max' figure for this chip (support article 102027), a CPU+GPU stress number; inference typically draws less. Apple measured the 32-core/80-GPU, 512GB config.",
    "generation": "previous",
    "status": "Discontinued 2026-08-25. Launch price; Apple no longer sells it new, so treat as a refurbished/used reference.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://support.apple.com/en-us/122211",
      "https://support.apple.com/en-us/102027",
      "https://lowendmac.com/2025/mac-studio-early-2025/"
    ]
  },
  {
    "id": "mac-studio-m3-ultra-256",
    "family": "Mac Studio",
    "chip": "M3 Ultra",
    "chip_variant": "32-core CPU / 80-core GPU",
    "unified_memory_gb": 256,
    "memory_bandwidth_gbs": 819,
    "usable_memory_gb": 192.0,
    "price_usd": 7099,
    "idle_watts": 9,
    "load_watts": 270,
    "load_watts_status": "published",
    "load_watts_note": "Apple's published 'max' figure for this chip (support article 102027), a CPU+GPU stress number; inference typically draws less.",
    "generation": "previous",
    "status": "Discontinued 2026-08-25. Launch price; Apple no longer sells it new, so treat as a refurbished/used reference. 256GB required the 32-core chip.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://support.apple.com/en-us/122211",
      "https://support.apple.com/en-us/102027",
      "https://lowendmac.com/2025/mac-studio-early-2025/"
    ]
  },
  {
    "id": "mac-studio-m3-ultra-512",
    "family": "Mac Studio",
    "chip": "M3 Ultra",
    "chip_variant": "32-core CPU / 80-core GPU",
    "unified_memory_gb": 512,
    "memory_bandwidth_gbs": 819,
    "usable_memory_gb": 384.0,
    "price_usd": 9499,
    "idle_watts": 9,
    "load_watts": 270,
    "load_watts_status": "published",
    "load_watts_note": "Apple's published 'max' figure for this chip (support article 102027), a CPU+GPU stress number; inference typically draws less.",
    "generation": "previous",
    "status": "Discontinued 2026-08-25. Launch price; Apple no longer sells it new, so treat as a refurbished/used reference. 512GB option withdrawn March 2026.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.",
    "sources": [
      "https://support.apple.com/en-us/122211",
      "https://support.apple.com/en-us/102027",
      "https://lowendmac.com/2025/mac-studio-early-2025/"
    ]
  },
  {
    "id": "nvidia-dgx-spark-128",
    "family": "DGX Spark",
    "chip": "GB10 Grace Blackwell",
    "chip_variant": "Founders Edition, 4TB",
    "unified_memory_gb": 128,
    "memory_bandwidth_gbs": 273,
    "usable_memory_gb": 119.5,
    "price_usd": 4699,
    "idle_watts": 5,
    "load_watts": 150,
    "load_watts_status": "third_party_measured",
    "load_watts_note": "NVIDIA publishes no system draw (240 W supply, GB10 TDP 140 W). 150 W is a single third-party unit's peak under inference (kubesimplify); Jeff Geerling measured 122.8 W running gpt-oss-20b.",
    "generation": "current",
    "status": "Shipping. MSRP raised from $3,999 to $4,699 in February 2026.",
    "notes": "CUDA deviceQuery reports 122,570 MiB (about 119.7 GiB) visible to the GPU under DGX OS; treated as 119.5 GB usable. Partner boxes (ASUS, Dell, Lenovo, Acer) exist but none was verifiably cheaper and in stock when checked.",
    "sources": [
      "https://www.nvidia.com/en-us/products/workstations/dgx-spark/",
      "https://forums.developer.nvidia.com/t/2-23-2026-price-change-announcement/361713",
      "https://blog.kubesimplify.com/day-3-the-dgx-spark-unpacked-gb10-unified-memory-sm-121-and-the-one-reason-this-hardware-exists",
      "https://github.com/geerlingguy/ai-benchmarks"
    ],
    "estimate_efficiency_moe": 0.65,
    "estimate_note": "MoE decode on CUDA reaches ~65% of the bandwidth ceiling here (Qwen3 30B-A3B: 89 tok/s measured vs 136 ceiling), against ~30% on Apple Silicon."
  },
  {
    "id": "framework-desktop-395-128",
    "family": "Strix Halo",
    "chip": "Framework Desktop",
    "chip_variant": "Ryzen AI Max+ 395 · Radeon 8060S, 40 CU",
    "unified_memory_gb": 128,
    "memory_bandwidth_gbs": 256,
    "usable_memory_gb": 96,
    "price_usd": 3449,
    "idle_watts": null,
    "load_watts": 133,
    "load_watts_status": "third_party_measured",
    "load_watts_note": "Jeff Geerling's ai-benchmarks: Framework Desktop mainboard at 133 W running Llama 3.1 70B in Ollama (98 W on gpt-oss-120b via Vulkan). Meter method not stated.",
    "generation": "current",
    "status": "System $3,449 (out of stock when checked); mainboard alone $3,149 on pre-order. Launched at $1,999 in 2025; LPDDR5x prices rose since.",
    "notes": "Windows lets you dedicate up to 75% of RAM to the GPU (Variable Graphics Memory in AMD Adrenalin; Framework documents 96 GB on a 128 GB board). On Linux the amdgpu GTT pool can be raised to roughly 108-120 GB with ttm.pages_limit / amdgpu.gttsize boot parameters. The Windows figure is used here as the conservative one. Measured GPU bandwidth is ~212-215 GB/s of the 256 GB/s theoretical, so estimates here run a little high.",
    "sources": [
      "https://frame.work/products/desktop-diy-amd-aimax300/configuration/new",
      "https://frame.work/blog/framework-desktop-deep-dive-ryzen-ai-max",
      "https://frame.work/desktop?tab=machine-learning",
      "https://community.frame.work/t/updated-commands-to-increase-max-unified-memory-usage-on-framework-desktop-under-fedora-43/78460",
      "https://github.com/geerlingguy/ai-benchmarks"
    ],
    "estimate_efficiency_moe": 0.55,
    "estimate_note": "MoE decode on Strix Halo reaches ~50-65% of the bandwidth ceiling (Qwen3 30B-A3B: 66-86 tok/s measured vs 128 ceiling), against ~30% on Apple Silicon."
  },
  {
    "id": "framework-desktop-395-64",
    "family": "Strix Halo",
    "chip": "Framework Desktop",
    "chip_variant": "Ryzen AI Max+ 395 · Radeon 8060S, 40 CU",
    "unified_memory_gb": 64,
    "memory_bandwidth_gbs": 256,
    "usable_memory_gb": 48,
    "price_usd": 1959,
    "idle_watts": null,
    "load_watts": 133,
    "load_watts_status": "third_party_measured",
    "load_watts_note": "Jeff Geerling's ai-benchmarks: Framework Desktop mainboard at 133 W running Llama 3.1 70B in Ollama (98 W on gpt-oss-120b via Vulkan). Meter method not stated. Measured on the 128 GB board; same chip.",
    "generation": "current",
    "status": "System $1,959 (out of stock when checked); mainboard alone $1,659. Framework lists 48 GB dedicated to the GPU.",
    "notes": "Windows lets you dedicate up to 75% of RAM to the GPU (Variable Graphics Memory in AMD Adrenalin; Framework documents 96 GB on a 128 GB board). On Linux the amdgpu GTT pool can be raised to roughly 108-120 GB with ttm.pages_limit / amdgpu.gttsize boot parameters. The Windows figure is used here as the conservative one. Measured GPU bandwidth is ~212-215 GB/s of the 256 GB/s theoretical, so estimates here run a little high.",
    "sources": [
      "https://frame.work/products/desktop-diy-amd-aimax300/configuration/new",
      "https://frame.work/desktop?tab=machine-learning",
      "https://github.com/geerlingguy/ai-benchmarks"
    ],
    "estimate_efficiency_moe": 0.55,
    "estimate_note": "MoE decode on Strix Halo reaches ~50-65% of the bandwidth ceiling (Qwen3 30B-A3B: 66-86 tok/s measured vs 128 ceiling), against ~30% on Apple Silicon."
  },
  {
    "id": "framework-desktop-385-32",
    "family": "Strix Halo",
    "chip": "Framework Desktop",
    "chip_variant": "Ryzen AI Max 385 · Radeon 8050S, 32 CU",
    "unified_memory_gb": 32,
    "memory_bandwidth_gbs": 256,
    "usable_memory_gb": 24,
    "price_usd": 1269,
    "idle_watts": null,
    "load_watts": 133,
    "load_watts_status": "stand_in",
    "load_watts_note": "No measurement for the Max 385 board; stand-in is the Max+ 395 board's measured 133 W, which likely overstates this smaller GPU.",
    "generation": "current",
    "status": "System $1,269 (out of stock when checked); mainboard alone $969. Note: the 32 GB tier is the smaller Ryzen AI Max 385 with a 32-CU GPU.",
    "notes": "Windows lets you dedicate up to 75% of RAM to the GPU (Variable Graphics Memory in AMD Adrenalin; Framework documents 96 GB on a 128 GB board). On Linux the amdgpu GTT pool can be raised to roughly 108-120 GB with ttm.pages_limit / amdgpu.gttsize boot parameters. The Windows figure is used here as the conservative one. Measured GPU bandwidth is ~212-215 GB/s of the 256 GB/s theoretical, so estimates here run a little high.",
    "sources": [
      "https://frame.work/products/desktop-diy-amd-aimax300/configuration/new",
      "https://frame.work/desktop?tab=machine-learning"
    ],
    "estimate_efficiency_moe": 0.55,
    "estimate_note": "MoE decode on Strix Halo reaches ~50-65% of the bandwidth ceiling (Qwen3 30B-A3B: 66-86 tok/s measured vs 128 ceiling), against ~30% on Apple Silicon."
  },
  {
    "id": "gmktec-evo-x2-128",
    "family": "Strix Halo",
    "chip": "GMKtec EVO-X2",
    "chip_variant": "Ryzen AI Max+ 395 · Radeon 8060S, 40 CU",
    "unified_memory_gb": 128,
    "memory_bandwidth_gbs": 256,
    "usable_memory_gb": 96,
    "price_usd": 3500,
    "idle_watts": null,
    "load_watts": 150,
    "load_watts_status": "third_party_measured",
    "load_watts_note": "ServeTheHome measured 147-160 W at the wall running Llama 3.3 70B Q6_K; 8-14 W idle.",
    "generation": "current",
    "status": "128 GB / 1 TB $3,499.99 in stock when checked (2 TB $3,649.99). Launched at about $2,000 in May 2025.",
    "notes": "Windows lets you dedicate up to 75% of RAM to the GPU (Variable Graphics Memory in AMD Adrenalin; Framework documents 96 GB on a 128 GB board). On Linux the amdgpu GTT pool can be raised to roughly 108-120 GB with ttm.pages_limit / amdgpu.gttsize boot parameters. The Windows figure is used here as the conservative one. Measured GPU bandwidth is ~212-215 GB/s of the 256 GB/s theoretical, so estimates here run a little high.",
    "sources": [
      "https://www.gmktec.com/products/amd-ryzen%e2%84%a2-ai-max-395-evo-x2-ai-mini-pc",
      "https://www.servethehome.com/gmktec-evo-x2-review-an-amd-ryzen-ai-max-395-powerhouse/4/"
    ],
    "estimate_efficiency_moe": 0.55,
    "estimate_note": "MoE decode on Strix Halo reaches ~50-65% of the bandwidth ceiling (Qwen3 30B-A3B: 66-86 tok/s measured vs 128 ceiling), against ~30% on Apple Silicon."
  },
  {
    "id": "gmktec-evo-x2-64",
    "family": "Strix Halo",
    "chip": "GMKtec EVO-X2",
    "chip_variant": "Ryzen AI Max+ 395 · Radeon 8060S, 40 CU",
    "unified_memory_gb": 64,
    "memory_bandwidth_gbs": 256,
    "usable_memory_gb": 48,
    "price_usd": 2200,
    "idle_watts": null,
    "load_watts": 150,
    "load_watts_status": "third_party_measured",
    "load_watts_note": "ServeTheHome's 147-160 W figure was on the 128 GB unit; same chip.",
    "generation": "current",
    "status": "64 GB / 1 TB $2,199.99 in stock when checked. 48 GB usable follows the same 75% Windows split Framework documents for a 64 GB board.",
    "notes": "Windows lets you dedicate up to 75% of RAM to the GPU (Variable Graphics Memory in AMD Adrenalin; Framework documents 96 GB on a 128 GB board). On Linux the amdgpu GTT pool can be raised to roughly 108-120 GB with ttm.pages_limit / amdgpu.gttsize boot parameters. The Windows figure is used here as the conservative one. Measured GPU bandwidth is ~212-215 GB/s of the 256 GB/s theoretical, so estimates here run a little high.",
    "sources": [
      "https://www.gmktec.com/products/amd-ryzen%e2%84%a2-ai-max-395-evo-x2-ai-mini-pc",
      "https://www.servethehome.com/gmktec-evo-x2-review-an-amd-ryzen-ai-max-395-powerhouse/4/",
      "https://frame.work/desktop?tab=machine-learning"
    ],
    "estimate_efficiency_moe": 0.55,
    "estimate_note": "MoE decode on Strix Halo reaches ~50-65% of the bandwidth ceiling (Qwen3 30B-A3B: 66-86 tok/s measured vs 128 ceiling), against ~30% on Apple Silicon."
  },
  {
    "id": "beelink-gtr9-pro-128",
    "family": "Strix Halo",
    "chip": "Beelink GTR9 Pro",
    "chip_variant": "Ryzen AI Max+ 395 · Radeon 8060S, 40 CU",
    "unified_memory_gb": 128,
    "memory_bandwidth_gbs": 256,
    "usable_memory_gb": 96,
    "price_usd": 4349,
    "idle_watts": null,
    "load_watts": 133,
    "load_watts_status": "stand_in",
    "load_watts_note": "No wall measurement found for this unit; stand-in is the Framework Desktop's measured 133 W on the same chip at the same 120 W sustained TDP.",
    "generation": "current",
    "status": "Only sold as 128 GB / 2 TB, $4,349 when checked (launched at $1,985).",
    "notes": "Windows lets you dedicate up to 75% of RAM to the GPU (Variable Graphics Memory in AMD Adrenalin; Framework documents 96 GB on a 128 GB board). On Linux the amdgpu GTT pool can be raised to roughly 108-120 GB with ttm.pages_limit / amdgpu.gttsize boot parameters. The Windows figure is used here as the conservative one. Measured GPU bandwidth is ~212-215 GB/s of the 256 GB/s theoretical, so estimates here run a little high.",
    "sources": [
      "https://www.bee-link.com/products/beelink-gtr9-pro-amd-ryzen-ai-max-395",
      "https://github.com/hogeheer499-commits/strix-halo-guide"
    ],
    "estimate_efficiency_moe": 0.55,
    "estimate_note": "MoE decode on Strix Halo reaches ~50-65% of the bandwidth ceiling (Qwen3 30B-A3B: 66-86 tok/s measured vs 128 ceiling), against ~30% on Apple Silicon."
  },
  {
    "id": "minisforum-ms-s1-max-128",
    "family": "Strix Halo",
    "chip": "Minisforum MS-S1 Max",
    "chip_variant": "Ryzen AI Max+ 395 · Radeon 8060S, 40 CU",
    "unified_memory_gb": 128,
    "memory_bandwidth_gbs": 256,
    "usable_memory_gb": 96,
    "price_usd": 3799,
    "idle_watts": null,
    "load_watts": 133,
    "load_watts_status": "stand_in",
    "load_watts_note": "No wall measurement found for this unit; stand-in is the Framework Desktop's measured 133 W on the same chip. Minisforum's performance mode allows 130 W.",
    "generation": "current",
    "status": "128 GB / 2 TB $3,799 on sale when checked (regular $4,749), shipping mid-September.",
    "notes": "Windows lets you dedicate up to 75% of RAM to the GPU (Variable Graphics Memory in AMD Adrenalin; Framework documents 96 GB on a 128 GB board). On Linux the amdgpu GTT pool can be raised to roughly 108-120 GB with ttm.pages_limit / amdgpu.gttsize boot parameters. The Windows figure is used here as the conservative one. Measured GPU bandwidth is ~212-215 GB/s of the 256 GB/s theoretical, so estimates here run a little high.",
    "sources": [
      "https://store.minisforum.com/products/minisforum-ms-s1-max-mini-pc"
    ],
    "estimate_efficiency_moe": 0.55,
    "estimate_note": "MoE decode on Strix Halo reaches ~50-65% of the bandwidth ceiling (Qwen3 30B-A3B: 66-86 tok/s measured vs 128 ceiling), against ~30% on Apple Silicon."
  },
  {
    "id": "hp-z2-mini-g1a-128",
    "family": "Strix Halo",
    "chip": "HP Z2 Mini G1a",
    "chip_variant": "Ryzen AI Max+ PRO 395 · Radeon 8060S, 40 CU",
    "unified_memory_gb": 128,
    "memory_bandwidth_gbs": 256,
    "usable_memory_gb": 96,
    "price_usd": 5544,
    "idle_watts": null,
    "load_watts": 133,
    "load_watts_status": "stand_in",
    "load_watts_note": "StorageReview saw the GPU alone drawing just under 100 W on gpt-oss-120b (GPU telemetry, not the wall); stand-in is the Framework Desktop's measured 133 W on the same silicon.",
    "generation": "current",
    "status": "128 GB / 1 TB $5,543.92 at Staples when checked; StorageReview cited a $4,781 list price in 2025. Enterprise pricing varies a lot by reseller.",
    "notes": "Windows lets you dedicate up to 75% of RAM to the GPU (Variable Graphics Memory in AMD Adrenalin; Framework documents 96 GB on a 128 GB board). On Linux the amdgpu GTT pool can be raised to roughly 108-120 GB with ttm.pages_limit / amdgpu.gttsize boot parameters. The Windows figure is used here as the conservative one. Measured GPU bandwidth is ~212-215 GB/s of the 256 GB/s theoretical, so estimates here run a little high.",
    "sources": [
      "https://www.staples.com/hp-z2-mini-g1a-desktop-computer-ryzen-ai-max-pro-395-128gb-ram-1tb-ssd-windows-11-pro-mouse-keyboard-included/product_IM1YS5467",
      "https://www.storagereview.com/review/hp-z2-mini-g1a-review-running-gpt-oss-120b-without-a-discrete-gpu"
    ],
    "estimate_efficiency_moe": 0.55,
    "estimate_note": "MoE decode on Strix Halo reaches ~50-65% of the bandwidth ceiling (Qwen3 30B-A3B: 66-86 tok/s measured vs 128 ceiling), against ~30% on Apple Silicon."
  },
  {
    "id": "macbook-air-m5-16",
    "family": "MacBook Air",
    "chip": "M5 (13-inch)",
    "chip_variant": "10-core CPU / 10-core GPU",
    "unified_memory_gb": 16,
    "memory_bandwidth_gbs": 153,
    "usable_memory_gb": 10.5,
    "price_usd": 1299,
    "idle_watts": null,
    "load_watts": 65,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has published no power figures for the 2026 Macs, and none at all for sustained GPU load on a laptop. Stand-in: the desktop maximum for the equivalent chip (M4 Mac mini, 65 W). A laptop draws less than that, so the electricity line here is an over-estimate.",
    "generation": "current",
    "status": "Shipping. Fanless, so it throttles hardest of any Mac here under a long generation.",
    "notes": "macOS lets the GPU wire roughly two-thirds of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. On a laptop, sustained speed drops once the chassis warms up and the fans cap out, so a desktop with the same chip will hold a higher tokens/sec over a long generation than the figures here suggest.",
    "sources": [
      "https://www.apple.com/macbook-air/specs/"
    ]
  },
  {
    "id": "macbook-pro-14-m5-16",
    "family": "MacBook Pro",
    "chip": "M5 (14-inch)",
    "chip_variant": "10-core CPU / 10-core GPU",
    "unified_memory_gb": 16,
    "memory_bandwidth_gbs": 153,
    "usable_memory_gb": 10.5,
    "price_usd": 1999,
    "idle_watts": null,
    "load_watts": 65,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has published no power figures for the 2026 Macs, and none at all for sustained GPU load on a laptop. Stand-in: the desktop maximum for the equivalent chip (M4 Mac mini, 65 W). A laptop draws less than that, so the electricity line here is an over-estimate.",
    "generation": "current",
    "status": "Shipping.",
    "notes": "macOS lets the GPU wire roughly two-thirds of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. On a laptop, sustained speed drops once the chassis warms up and the fans cap out, so a desktop with the same chip will hold a higher tokens/sec over a long generation than the figures here suggest.",
    "sources": [
      "https://www.apple.com/macbook-pro/specs/",
      "https://www.apple.com/shop/buy-mac/macbook-pro",
      "https://support.apple.com/en-us/102059"
    ]
  },
  {
    "id": "macbook-pro-14-m5-32",
    "family": "MacBook Pro",
    "chip": "M5 (14-inch)",
    "chip_variant": "10-core CPU / 10-core GPU",
    "unified_memory_gb": 32,
    "memory_bandwidth_gbs": 153,
    "usable_memory_gb": 21.0,
    "price_usd": 2399,
    "idle_watts": null,
    "load_watts": 65,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has published no power figures for the 2026 Macs, and none at all for sustained GPU load on a laptop. Stand-in: the desktop maximum for the equivalent chip (M4 Mac mini, 65 W). A laptop draws less than that, so the electricity line here is an over-estimate.",
    "generation": "current",
    "status": "Shipping.",
    "notes": "macOS lets the GPU wire roughly two-thirds of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. On a laptop, sustained speed drops once the chassis warms up and the fans cap out, so a desktop with the same chip will hold a higher tokens/sec over a long generation than the figures here suggest.",
    "sources": [
      "https://www.apple.com/macbook-pro/specs/",
      "https://www.apple.com/shop/buy-mac/macbook-pro",
      "https://support.apple.com/en-us/102059"
    ]
  },
  {
    "id": "macbook-pro-16-m5-pro-24",
    "family": "MacBook Pro",
    "chip": "M5 Pro (16-inch)",
    "chip_variant": "15-core CPU / 16-core GPU",
    "unified_memory_gb": 24,
    "memory_bandwidth_gbs": 307,
    "usable_memory_gb": 16.0,
    "price_usd": 2999,
    "idle_watts": null,
    "load_watts": 140,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has published no power figures for the 2026 Macs, and none at all for sustained GPU load on a laptop. Stand-in: the desktop maximum for the equivalent chip (M4 Pro Mac mini, 140 W). A laptop draws less than that, so the electricity line here is an over-estimate.",
    "generation": "current",
    "status": "Shipping.",
    "notes": "macOS lets the GPU wire roughly two-thirds of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. On a laptop, sustained speed drops once the chassis warms up and the fans cap out, so a desktop with the same chip will hold a higher tokens/sec over a long generation than the figures here suggest.",
    "sources": [
      "https://www.apple.com/macbook-pro/specs/",
      "https://www.apple.com/shop/buy-mac/macbook-pro",
      "https://support.apple.com/en-us/102059"
    ]
  },
  {
    "id": "macbook-pro-16-m5-pro-48",
    "family": "MacBook Pro",
    "chip": "M5 Pro (16-inch)",
    "chip_variant": "15-core CPU / 16-core GPU",
    "unified_memory_gb": 48,
    "memory_bandwidth_gbs": 307,
    "usable_memory_gb": 36.0,
    "price_usd": 3599,
    "idle_watts": null,
    "load_watts": 140,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has published no power figures for the 2026 Macs, and none at all for sustained GPU load on a laptop. Stand-in: the desktop maximum for the equivalent chip (M4 Pro Mac mini, 140 W). A laptop draws less than that, so the electricity line here is an over-estimate.",
    "generation": "current",
    "status": "Shipping.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. On a laptop, sustained speed drops once the chassis warms up and the fans cap out, so a desktop with the same chip will hold a higher tokens/sec over a long generation than the figures here suggest.",
    "sources": [
      "https://www.apple.com/macbook-pro/specs/",
      "https://www.apple.com/shop/buy-mac/macbook-pro",
      "https://support.apple.com/en-us/102059"
    ]
  },
  {
    "id": "macbook-pro-16-m5-pro-64",
    "family": "MacBook Pro",
    "chip": "M5 Pro (16-inch)",
    "chip_variant": "15-core CPU / 16-core GPU",
    "unified_memory_gb": 64,
    "memory_bandwidth_gbs": 307,
    "usable_memory_gb": 48.0,
    "price_usd": 3999,
    "idle_watts": null,
    "load_watts": 140,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has published no power figures for the 2026 Macs, and none at all for sustained GPU load on a laptop. Stand-in: the desktop maximum for the equivalent chip (M4 Pro Mac mini, 140 W). A laptop draws less than that, so the electricity line here is an over-estimate.",
    "generation": "current",
    "status": "Shipping.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. On a laptop, sustained speed drops once the chassis warms up and the fans cap out, so a desktop with the same chip will hold a higher tokens/sec over a long generation than the figures here suggest.",
    "sources": [
      "https://www.apple.com/macbook-pro/specs/",
      "https://www.apple.com/shop/buy-mac/macbook-pro",
      "https://support.apple.com/en-us/102059"
    ]
  },
  {
    "id": "macbook-pro-16-m5-max-48",
    "family": "MacBook Pro",
    "chip": "M5 Max (16-inch)",
    "chip_variant": "18-core CPU / 40-core GPU",
    "unified_memory_gb": 48,
    "memory_bandwidth_gbs": 614,
    "usable_memory_gb": 36.0,
    "price_usd": 4999,
    "idle_watts": null,
    "load_watts": 145,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has published no power figures for the 2026 Macs, and none at all for sustained GPU load on a laptop. Stand-in: the desktop maximum for the equivalent chip (M4 Max Mac Studio, 145 W). A laptop draws less than that, so the electricity line here is an over-estimate.",
    "generation": "current",
    "status": "Shipping.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. On a laptop, sustained speed drops once the chassis warms up and the fans cap out, so a desktop with the same chip will hold a higher tokens/sec over a long generation than the figures here suggest.",
    "sources": [
      "https://www.apple.com/macbook-pro/specs/",
      "https://www.apple.com/shop/buy-mac/macbook-pro",
      "https://support.apple.com/en-us/102059"
    ]
  },
  {
    "id": "macbook-pro-16-m5-max-64",
    "family": "MacBook Pro",
    "chip": "M5 Max (16-inch)",
    "chip_variant": "18-core CPU / 40-core GPU",
    "unified_memory_gb": 64,
    "memory_bandwidth_gbs": 614,
    "usable_memory_gb": 48.0,
    "price_usd": 5399,
    "idle_watts": null,
    "load_watts": 145,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has published no power figures for the 2026 Macs, and none at all for sustained GPU load on a laptop. Stand-in: the desktop maximum for the equivalent chip (M4 Max Mac Studio, 145 W). A laptop draws less than that, so the electricity line here is an over-estimate.",
    "generation": "current",
    "status": "Shipping.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. On a laptop, sustained speed drops once the chassis warms up and the fans cap out, so a desktop with the same chip will hold a higher tokens/sec over a long generation than the figures here suggest.",
    "sources": [
      "https://www.apple.com/macbook-pro/specs/",
      "https://www.apple.com/shop/buy-mac/macbook-pro",
      "https://support.apple.com/en-us/102059"
    ]
  },
  {
    "id": "macbook-pro-16-m5-max-128",
    "family": "MacBook Pro",
    "chip": "M5 Max (16-inch)",
    "chip_variant": "18-core CPU / 40-core GPU",
    "unified_memory_gb": 128,
    "memory_bandwidth_gbs": 614,
    "usable_memory_gb": 96.0,
    "price_usd": 6999,
    "idle_watts": null,
    "load_watts": 145,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has published no power figures for the 2026 Macs, and none at all for sustained GPU load on a laptop. Stand-in: the desktop maximum for the equivalent chip (M4 Max Mac Studio, 145 W). A laptop draws less than that, so the electricity line here is an over-estimate.",
    "generation": "current",
    "status": "Shipping.",
    "notes": "macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. On a laptop, sustained speed drops once the chassis warms up and the fans cap out, so a desktop with the same chip will hold a higher tokens/sec over a long generation than the figures here suggest.",
    "sources": [
      "https://www.apple.com/macbook-pro/specs/",
      "https://www.apple.com/shop/buy-mac/macbook-pro",
      "https://support.apple.com/en-us/102059"
    ]
  },
  {
    "id": "gmktec-evo-x3-128",
    "family": "Strix Halo",
    "chip": "GMKtec EVO-X3",
    "chip_variant": "Ryzen AI Max+ 395 · Radeon 8060S, 40 CU",
    "unified_memory_gb": 128,
    "memory_bandwidth_gbs": 256,
    "usable_memory_gb": 96,
    "price_usd": 3600,
    "idle_watts": null,
    "load_watts": 150,
    "load_watts_status": "third_party_measured",
    "load_watts_note": "No separate measurement for the X3; ServeTheHome measured 147-160 W at the wall on the EVO-X2, same chip and TDP.",
    "generation": "current",
    "status": "Shipping. Replaced the EVO-X2 in July 2026.",
    "notes": "Windows lets you dedicate up to 75% of RAM to the GPU (Variable Graphics Memory in AMD Adrenalin). On Linux the amdgpu GTT pool can be raised to roughly 108-120 GB with ttm.pages_limit / amdgpu.gttsize boot parameters. The Windows figure is used here as the conservative one. Measured GPU bandwidth is ~212-215 GB/s of the 256 GB/s theoretical, so estimates here run a little high.",
    "sources": [
      "https://www.gmktec.com/products/amd-ryzen-ai-max-395-evo-x3-ai-mini-pc"
    ],
    "estimate_efficiency_moe": 0.55,
    "estimate_note": "MoE decode on Strix Halo reaches ~50-65% of the bandwidth ceiling (Qwen3 30B-A3B: 66-86 tok/s measured vs 128 ceiling), against ~30% on Apple Silicon."
  },
  {
    "id": "framework-desktop-495-192",
    "family": "Strix Halo",
    "chip": "Framework Desktop",
    "chip_variant": "Ryzen AI Max+ PRO 495 'Gorgon Halo' · 192 GB",
    "unified_memory_gb": 192,
    "memory_bandwidth_gbs": 273,
    "usable_memory_gb": 160,
    "price_usd": null,
    "idle_watts": null,
    "load_watts": 133,
    "load_watts_status": "stand_in",
    "load_watts_note": "No measurement for Gorgon Halo; stand-in is the Framework Desktop's measured 133 W on the previous chip.",
    "generation": "current",
    "status": "Announced, not yet priced or orderable.",
    "notes": "192 GB unified, of which AMD states 160 GB can be allocated to the GPU — the largest GPU-addressable pool of anything on this list under $10,000, if it ships at a sane price. Windows lets you dedicate up to 75% of RAM to the GPU (Variable Graphics Memory in AMD Adrenalin). On Linux the amdgpu GTT pool can be raised to roughly 108-120 GB with ttm.pages_limit / amdgpu.gttsize boot parameters. The Windows figure is used here as the conservative one. Measured GPU bandwidth is ~212-215 GB/s of the 256 GB/s theoretical, so estimates here run a little high.",
    "sources": [
      "https://frame.work/desktop"
    ],
    "estimate_efficiency_moe": 0.55,
    "estimate_note": "MoE decode on Strix Halo reaches ~50-65% of the bandwidth ceiling (Qwen3 30B-A3B: 66-86 tok/s measured vs 128 ceiling), against ~30% on Apple Silicon.",
    "TODO": "price_usd: Framework has not announced the 192 GB price."
  },
  {
    "id": "macbook-air-m5-15-16",
    "family": "MacBook Air",
    "chip": "M5 (15-inch)",
    "chip_variant": "10-core CPU / 10-core GPU",
    "unified_memory_gb": 16,
    "memory_bandwidth_gbs": 153,
    "usable_memory_gb": 10.5,
    "price_usd": 1499,
    "idle_watts": null,
    "load_watts": 65,
    "load_watts_status": "stand_in",
    "load_watts_note": "Apple has published no power figures for the 2026 Macs, and none at all for sustained GPU load on a laptop. Stand-in: the desktop maximum for the equivalent chip (M4 Mac mini, 65 W). A laptop draws less than that, so the electricity line here is an over-estimate.",
    "generation": "current",
    "status": "Shipping. Fanless, so it throttles hardest of any Mac here under a long generation.",
    "notes": "macOS lets the GPU wire roughly two-thirds of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. On a laptop, sustained speed drops once the chassis warms up and the fans cap out, so a desktop with the same chip will hold a higher tokens/sec over a long generation than the figures here suggest.",
    "sources": [
      "https://www.apple.com/macbook-air/specs/"
    ]
  }
]