MiniPC Guru @MiniPCGuru

@paulund For always-on coding agents, an N150 mini PC idles around 6W, so it runs 24/7 without the power bill noticing.

Lowest tracked price plus full specs: minipc.guru/mini-pc/minisf…

MINISFORUM UM790 Pro with the Ryzen 9 7940HS, 32GB RAM and 1TB SSD is down to $655.02 at NextWarehouse, in stock. Same config sits at $869 on Amazon, so that's about $214 less. Two USB4 ports and dual M.2 slots at that price make it a lot of box for a homelab or 1080p gaming. Price last checked yesterday, so it can move.

@RaidOwlTweets For apples-to-apples between boxes, run everything through the same llama.cpp server build with the same quant. Swapping harnesses between runs tells you more about the tool than the hardware.

minipc.guru/compare?items=…

Beelink SER8 at $939 vs MINISFORUM UM790 Pro at $655. Same 32GB/1TB, same Radeon 780M, both Zen 4 — and the UM790 Pro even has two USB4 ports. Buy the SER8 if you want the newer 8845HS and its beefier NPU. Everyone else should keep the $284 and take the UM790 Pro.

@Acemagic_ Worth noting the 13900F ships without an iGPU, so that 191.9 TFLOPS figure has to come from the discrete acceleration in the box. Curious what real token/s a 13B hits on it.

@45Homelab Four tool-free bays and an Unraid license in the box is exactly how you win the starter-homelab crowd. Idle power draw will decide whether it earns a permanent shelf spot though.

@gmktecdotcom 192GB of unified memory on a Strix Halo-class APU is what makes this box interesting — that's the spec that puts 120B-class models in RAM on a desktop footprint.

@Hi_MINISFORUM Strix Halo is the interesting part here for local AI: unified memory lets the iGPU claim a big share of system RAM, so much bigger models fit on one box. If you're buying it for local LLMs, the RAM config is the spec that decides which models fit.

@geekompc The A5 is the quiet one: whisper-quiet, fits 3 drives. Pick it when the box lives in the room where you sleep.

@Acemagic_ The number that matters for local AI on this box: 128GB unified memory, up to 96GB allocatable to the GPU. That's a 70B at Q4 on one mini PC.

Full side-by-side specs, prices, and noise numbers: minipc.guru/compare?items=…

Two $450 boxes, two different jobs. GEEKOM A5 (7430U, 1TB): whisper-quiet, fits 3 drives. The silent desk PC / light Proxmox host. GMKtec M5 Ultra (7730U, 512GB): dual 2.5GbE. The firewall box. Silence → A5. Networking → M5 Ultra.

Deal via PCWorld's early Prime Day roundup — Prime Big Deal Days run Oct 6-7, this one is live now: pcworld.com/article/324451…

PCWorld named the Acemagic M5 its top mini PC deal ahead of October Prime Day: $563.97 (37% off) — Intel Core 5, 24GB RAM, 512GB SSD. My take: upgradeable RAM/storage and triple-display support make it a good office/homelab box at this price. Just don't buy it to game.

@Hi_MINISFORUM Flash-Next is the easy warmup for the MS-S1 MAX — the real test is a 70B at Q4 sustained. The P495 SKU's spec that matters most: up to 192GB of unified memory, with the full pool usable for weights instead of being capped at dGPU VRAM.

Pre-orders open tomorrow morning PT. For local LLMs, the number that matters: 192GB unified memory, up to 160GB for the GPU. 128GB Strix Halo boxes cap GPU memory at 96GB — this raises the ceiling 50%. Framework demos DeepSeek V4-Flash at Q8 on one box. No pricing yet.

Framework @FrameworkPuter

2 weeks ago

We're opening pre-orders for the 192GB Framework Desktop this Wednesday morning Pacific Time. This has AMD's new Ryzen AI Max+ Pro 495, an open-end x4 PCIe slot, a pre-installed Noctua fan, and an optional pre-built config with Fedora pre-loaded.

202 136 3K 361K 441

The BD395i MAX is the board that makes Strix Halo interesting for homelabs: Mini-ITX with the full Ryzen AI Max+ 395 and up to 128GB of unified memory, which is the part that actually lets you run 70B-class models on a board this size. A serious board-level cooling design is exactly what the laptop-first 395 boxes have been missing.

This is the whole decision tree in one line. A 70B at Q4 needs ~40GB, at Q8 ~70GB, so both fit comfortably in a 128GB Strix Halo box with room for big context. The 192GB tier only earns its ~2x premium if you're running 120B+ at high quants or serving huge context windows. Everyone else should pocket the difference.