$4K AMD Strix Halo Mini PC with 128GB Unified Memory Runs Full-Repo Local LLM Workflows
MooseEfficient2151 · reddit · 2026-10-20
A hands-on Reddit report building on TechPowerUp's review: AMD's Ryzen AI Max+ 395 (Strix Halo) mini PC costs $4K and packs 128GB of unified memory, letting it run local models that 32GB GPUs can't fit.
Versus NVIDIA's DGX Spark: Spark has faster time-to-first-token, CUDA, and fp8/fp4 support, but the 128GB version costs $7K. Strix Halo offers Windows/Linux support and a fully private, subscription-free local setup, at the cost of lower memory bandwidth, idle fan noise, and missing advanced precision formats.
The poster's real-world test: the old bottleneck was blowing past VRAM when dumping whole codebases into the KV cache. On the 128GB box they loaded qwen3.8-flash and ran multi-file refactoring with codex piped through sumus—model and entire repo coexisting in unified memory, fully offline.
More from Infra
- Google Cloud london region outage hits dozens of services; Revolut and Lloyds among affected — neal_lathia · 2026-10-20
- NVIDIA's $500B AI infra deal had "no obvious provenance" as PE debt talks stall — mjdramstead · 2026-10-20
- The Local Computing Rebellion: Subscription Fatigue and Forced AI Drive a Shift Back On-Premises — DavidLinthicum · 2026-10-20
- DeepSeek also open-sources deepseek-recipe, replicating official API token encoding logic — deepseek_ai · 2026-10-20
- DeepSeek open-sources DeepInfer, the production inference framework behind DeepSeek-V4.1-Flash — deepseek_ai · 2026-10-20
- Scott Hanselman serves local inference to his whole house from a Surface Laptop Ultra — lee_stott · 2026-10-20