STRIX HALO VS MAC STUDIO

answer: For big models per dollar, a 128GB Strix Halo box; for speed per token, a Mac Studio with enough memory — but the starting Mac Studio M5 Max has only 36GB, so it runs less than a 128GB Strix Halo box.
strix_halo_128gb: $3,449 · UP 73% SINCE FEB 2025
mac_studio_m5_max: $2,499 (36GB start) · 460 GB/s

HOW THIS IS DETERMINED

compare:
  - memory decides what fits (columns 70B / 120B)
  - bandwidth decides tokens per second once it fits
  - price and price trend decide when to buy
prices: official maker list / MSRP only, latest recorded point, linked in SOURCES
fit: weights file size (Hugging Face) + KV cache at 8,192 tokens (model config.json) + 10% overhead, compared with usable memory: VRAM − 1 GB, or 75% of unified / system memory (details)
speed: only published llama.cpp runs; we never estimate tok/s
not_considered: street prices, used hardware, multi-GPU builds, power and noise

LIVE TABLE

Framework Desk 128GB128GB unified256 GB/s$3,449YESYES47.5 [1]UP 73% SINCE FEB 2025BUY NOW → Framework
AMD Ryzen AI Halo128GB unified256 GB/s$3,999.99 (Jul 11)YESYES47.5 [1]*NEW*BUY NOW → Micro Center
Corsair AI WS 300 128GB128GB unified256 GB/s$4,599.99 (Oct 29)YESYES47.5 [1]DOWN 2.1% SINCE AUG 2026BUY NOW → Corsair
GMKtec EVO-X2128GB unified256 GB/s$3,499.99 (Oct 15)YESYES47.5 [1]UP 6.1% SINCE MAY 2026BUY NOW → GMKtec
Minisforum MS-S1 MAX128GB unified256 GB/s$2,919.90 (Sep 13)YESYES47.5 [1]FLAT SINCE JAN 2026BUY NOW → Minisforum
DGX Spark 128GB128GB unified273 GB/s$6,950 (Oct 10)YESYES58.7 [2]UP 74% SINCE NOV 2025BUY NOW → NVIDIA
Mac Studio M5 Max36GB unified460 GB/s$2,499nono—FLAT SINCE SEP 2026BUY NOW → Apple
Mac Studio M5 Ultra64GB unified1200 GB/s$5,499TIGHTno—FLAT SINCE SEP 2026BUY NOW → Apple
Mac mini M5 Pro24GB unified307 GB/s$1,699nono—FLAT SINCE SEP 2026BUY NOW → Apple

SOURCES

Current sources only. Every older price point and its source is in the paid export.

#whatsourcefetched
[1]GMKtec EVO-X2 (Ryzen AI Max+ 395, 128GB), ROCm, posted in llama.cpp #16578github.com/ggml-org/llama.cpp/discussions/165782026-11-13
[2]ggml-org DGX Spark bench, build 7941, depth 0github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spa...2026-11-13