TokenAssemble

Browse · 14 tracked

Systems for local LLMs

We track 14 prebuilt systems for local LLM inference — 7 Apple-silicon Macs plus the DGX Spark and Strix Halo mini-PC class — with sourced memory, bandwidth, and as-configured prices, each graded against every model we cover. Unified memory is shared with the OS — the usable figure is what the GPU can actually address. “Runs up to” is the largest model that loads cleanly (grade B or better) with llama.cpp at 8K context — see how the math works.

Apple silicon

SystemMemoryBandwidthPriceRuns up to
Mac Studio M3 Ultra 96GB96 GB (~72 usable)819 GB/s$5,299Qwen2.5-72B-Instruct
Mac Studio M4 Max 64GB64 GB (~48 usable)546 GB/s$3,499Qwen3.6-35B-A3B
Mac mini M4 Pro 48GB48 GB (~36 usable)273 GB/s$2,199Qwen3.6-35B-A3B
Mac Studio M4 Max 36GB36 GB (~27 usable)410 GB/s$2,499Qwen3.6-35B-A3B
Mac mini M4 24GB24 GB (~18 usable)120 GB/s$999gpt-oss-20b
Mac mini M4 Pro 24GB24 GB (~18 usable)273 GB/s$1,599gpt-oss-20b
Mac mini M4 16GB16 GB (~12 usable)120 GB/s$799Mistral-Nemo-Instruct-2407

Mini-PCs & AI workstationsbeta tiers

SystemMemoryBandwidthPriceRuns up to
ASUS Ascent GX10128 GB (~115 usable)273 GB/s$3,999DeepSeek-V4-Flash
NVIDIA DGX Spark128 GB (~115 usable)273 GB/s$4,699DeepSeek-V4-Flash
Beelink GTR9 Pro Ryzen AI Max+ 395 128GB128 GB (~96 usable)256 GB/s$4,349gpt-oss-120b
Framework Desktop Ryzen AI Max+ 395 128GB128 GB (~96 usable)256 GB/s$3,449gpt-oss-120b
GMKtec EVO-X2 Ryzen AI Max+ 395 128GB128 GB (~96 usable)256 GB/s$3,399.99gpt-oss-120b
HP Z2 Mini G1a Ryzen AI Max+ PRO 395 128GB128 GB (~96 usable)256 GB/s$5,021gpt-oss-120b
Framework Desktop Ryzen AI Max+ 395 64GB64 GB (~48 usable)256 GB/s$1,959Qwen3.6-35B-A3B

Fit is physics and holds everywhere; mini-PC-class speed tiers are beta until real measurements calibrate them. Prices are as-configured and stamped on each system's page.