Good morning — it's Monday, November 2, 2026. The lead: ASRock's Strix Halo box with a carry handle, the fastest of its kind ServeTheHome has tested. From the bench: Mage-VL-Cowork, a 4B model that clicks or codes, per turn. Plus a map of your whole lab, a Mattermost fork with no license key, and browser CAD you can host yourself. Don't keep us a secret: forward this to a fellow home labber, or share the web version.

Share this issue:  X   WhatsApp   Reddit   Facebook   LinkedIn   Text   Email

The fastest Strix Halo box ServeTheHome has tested wears a handle

Most Strix Halo mini PCs are black boxes. ASRock Industrial's AI Box-A395 is a small aluminum tower instead: a Ryzen AI Max+ 395 (16 cores, 5.1 GHz) with Radeon 8060S graphics and 128GB of soldered LPDDR5X-8000, two empty M.2 slots, a 400W internal power supply — no brick to haul — and a carry handle in the box. It weighs 2.8 kg.

ServeTheHome's testing (independent) calls it the fastest Ryzen AI Max+ 395 system it has reviewed: the chip ran at 130–140W under load, held 4.6–4.7 GHz for hours, and stayed near 35 dBA at idle and 42 dBA under load. The rear is well stocked — two 40 Gbps USB-C ports, 10GbE plus 2.5GbE — and the industrial extras are real: dual BIOS and a watchdog timer. The caveats are printed as plainly: the only top ports are two USB-A at 480 Mbps, so every fast connection lives at the back, and the M.2 slots are cramped. ServeTheHome lists no price, so none runs here. For your rack: the Strix Halo box to beat for an always-on local AI node — best run headless, not as a desk workstation.

Mage-VL-Cowork — a 4B agent that knows when to click and when to code

Microsoft's Mage-VL-Cowork, released October 30, puts two tools in one 4B vision-language policy: screenshot-grounded clicking, and a persistent terminal. Per turn it picks the channel — clicking through a visual layout, or editing the spreadsheet in Python instead of through forty menus. A useful shape for a lab agent, at a size a home GPU can hold.

Verdict first: on Microsoft's own numbers, the 4B model edges the 27B entry in the same table. Higher is better throughout.
OSWorld-Verified overall: Mage-VL-Cowork 66.0% — Qwen3.6-27B 63.9% — Mage ahead by 2.1.
OS tasks: Mage-VL-Cowork 87.5% — no Qwen breakdown published; Microsoft's harness only.
Multi-app workflows: Mage-VL-Cowork 62.7% — Microsoft's harness only.
Every score above is (vendor): Microsoft's card, Microsoft's harness, 361 tasks, no independent replication yet.

Setup notes: SGLang via Microsoft's pinned Docker image, or Transformers offline; the custom architecture needs trust_remote_code, and there is no Ollama listing. Difficulty: fiddly. The card's own warning: this model runs arbitrary shell commands by design, so run it only in an isolated, disposable sandbox.

GitHub find: Homelable draws the rack you actually have. This self-hosted visualiser scans your network, imports from Proxmox and Zigbee2MQTT, and draws rack canvases with port-to-port patch cables, live health checks, and a document per device. MIT-licensed, 4,871 stars, latest release v3.6.0 on October 6.

AntiMatter forks Mattermost and drops the license key. A self-hosted team chat server — channels, threads, file sharing, calls — where every feature is available without a license key, and the Mattermost apps still work. One Docker image runs it. It is young (the repo opened in September), so try it before you migrate to it.

Layerling is browser CAD for 3D printing that you can host. Fillets, chamfers, hollowing and threads, with export to STL, 3MF and STEP, no account, and designs that stay on your computer. It is AGPL-3.0 free software, and its Docker image is aimed squarely at NAS boxes and home servers.

Monday, 8:20 PM. Your core switch has started rebooting itself every twenty minutes, and the whole lab drops with it each time. On the shelf sit a spare unmanaged switch and last week's config backup — and two people in this house work from home through that switch tomorrow morning.

Your call:
A. Move the critical cables to the spare switch, restore service for the night, and bench-test the core switch offline
B. Move the core switch to a different outlet and UPS output and watch whether the reboots stop
C. Factory-reset the core switch in place and rebuild its configuration from memory tonight

100 Hero Points for the right call. Your score is on the outcome page, ready to share.

Reading this because a fellow home labber forwarded it? Get your own copy every morning — subscribe free, and pass it on — that is how this grows.