3 weeks with a Minisforum N5 Max (Strix Halo, 128 GB) as an all-in-one Proxmox NAS + media + SSO + local AI box — full write-up, gotchas, mistakes, and a 2026 value comparison
retained text
I built my household server on a \*\*Minisforum N5 Max\*\* (Ryzen AI Max+ 395, Radeon 8060S, 128 GB unified RAM, 5 bays) running \*\*Proxmox VE\*\*, and documented everything — including the mistakes: [https://github.com/Mensra/n5-max-homelab](https://github.com/Mensra/n5-max-homelab) \*\*What it runs:\*\* ZFS raidz2 (5×14 TB) with SMB shares and Time Machine; PBS backups to a separate 48 TB USB enclosure; Jellyfin with iGPU transcoding; Immich; TeslaMate; AdGuard DNS with automatic router failover; Caddy; Authentik SSO (OIDC + LDAP for Samba); plus Ollama (a 70B model fits), Open WebUI and ComfyUI running \*\*full-quality FLUX.2 dev\*\*, all offline. \*\*A few things I wish I'd known:\*\* \* The \*\*M.2 heatsink shipped loose in the accessory box\*\*, and only one M.2 slot is PCIe 4.0 x4. \* The \*\*HDD fans follow ambient temperature, not the drives\*\* — base fan speed is what actually cools the disks under load. \* \*\*GPU memory (GTT) isn't counted against any container's memory limit.\*\* A big model can starve the host while every limit looks fine. \* \*\*ComfyUI needs \`--disable-mmap\` on Strix Halo\*\*, or every model costs \~2x its size in RAM. \* \*\*Ubuntu's stock \`nftables\` service wipes Docker's firewall rules when the package updates\*\* — in your DNS container, that takes the house offline. \* \*\*Jellyfin: passing \`/dev/dri\` isn't enough.\*\* Mine transcoded on the CPU for two weeks before I noticed. \* \*\*I tested my SSO fallbacks with Authentik actually stopped\*\* and found documented passwords that didn't work anywhere. Test yours. There's also a [MISTAKES.md](http://MISTAKES.md) with the times I broke things (including an unquoted heredoc that ran a live \`zfs rollback\`). \*\*Is it worth it?\*\* With 128 GB of DDR5 alone around $3,400 this year, I couldn't find anything cheaper that matches it for storage + local AI, even as a multi-box setup. The GPU is mid-range in speed (a 1536² FLUX.2 image takes \~18 min), but it \*holds\* models no consumer card can. Comparison with sources in VALUE.md. Happy to answer questions.
document fetched 2026-09-27 00:34 · canonical https://reddit.com/r/homelab/comments/1wqs7ii/3_weeks_with_a_minisforum_n5_max_strix_halo_128