Local LLM hardware: ASUS GX10 vs Mac Mini M4 Pro
Quote188 words
Both machines have ~275 GB/s bandwidth but 128GB vs 24GB unified memory, so bandwidth sets token speed and memory sets what fits. Low-active MoE models in the 30-120B range are the sweet spot, while a 235B model crawls at ~1.5 tok/s on the GX10.
Key claims
Qwen3-30B-A3B runs at ~10-15 tok/s on the Mac Mini
GX10 fits models up to 235B at Q4
235B sits at ~1.5 tok/s due to the bandwidth ceiling
Time to build (with local LLMs).
Scaled my hardware setup with an ASUS GX10 and a Mac Mini M4 Pro.
Specs that matter:
- Bandwidth is basically identical. ~275 GB/s on both
- Memory is where things diverge: 24GB vs 128GB unified
TL;DR:
Bandwidth = how FAST tokens come out.
Memory = what FITS.
Hence, the GX10’s real edge here is size, not speed.
So, what LLMs can I actually run?
Mac Mini (~16-18GB usable after macOS):
- Comfy to 14B dense
- 30B MoE fits but tight
- Qwen3-30B-A3B ~18GB, 10-15 tok/s (only ~3B active)
- 70B+ no chance
GX10 (~120GB usable):
- Fits anything up to 235B at Q4
- But the bandwidth ceiling means high-active models crawl. 235B sits ~1.5 tok/s, hence cold standby
- Sweet spot = low-active MoE, 30-120B range (e.g. 30B-A3B or GLM-4.5-Air)
The rule for both:
Total parameters = does it fit.
Active parameters = how fast it runs.
That’s exactly why MoE is the play on this kind of hardware.
Tbh tho, will likely use only GX10 for running LLMs and make the mac mini the cockpit / general workhorse.
QUOTING @zerokn0wledge_
Claude Fable 5, the most capable model ever released to the public, lasted three days before a US export-control order pulled it offline. But you know the story. That's why I'm here to tell you about the part that REALLY matters: Anything you "rent" can be switched off by som
Claude Fable 5, the most capable model ever released to the public, lasted three days before a US export-control order pulled it offline. But you know the story. That's why I'm here to tell you about the part that REALLY matters: Anything you "rent" can be switched off by som
1 image on X