Good morning — it's Friday, November 6. Supermicro's Super AI Station is a six-digit AI server that converts from desk tower to 5U rackmount. From the bench: GLB-3D-Coder, a 4B model that writes exportable 3D assets. Drops: Open WebUI v0.13.0, Immich's telemetry fix, and a 280B Baidu base. Forward this to a fellow home labber.
Share this issue: X WhatsApp Reddit Facebook LinkedIn Text Email
Supermicro's Super AI Station, reviewed by ServeTheHome yesterday, takes the server-at-your-desk idea literally: the same chassis stands as a full-size tower or slides onto rails as a 5U rackmount server. At 88 pounds and 28 inches deep, it is a tower in name only.
Inside: a 72-core Grace CPU with 496GB of LPDDR5X, a Blackwell Ultra GPU with 252GB of HBM3E, and two 400Gbps ConnectX-8 ports that can push 800Gbps combined. In ServeTheHome's runs (independent), several tool-call scenarios on Qwen3.8 Flash Next passed 3,000 tokens per second at 32-way concurrency — an agent swarm on one box.
The surprise is the manners. Its 1,600W Titanium supply never idles below 250W, yet at full load the liquid-cooled box measured 39–42 dBA — the team kept not noticing it was running. The price is six digits — not a buying tip, but a marker: the lab's aspirational machine now converts between desk and rack.
CCORE GLB-3D-Coder, released yesterday by independent developer CCORE from a Qwen3.5-4B base, does one narrow thing: describe a character, prop or building, and it writes a complete Three.js page with a preview and an Export GLB button, ready for Godot. A playable browser arena, RIFTWARD, was built from its assets, and its training dataset shipped the same day: 3,918 examples, 40 domains.
On the evidence, honestly: the card publishes no numeric scores, so there is no scoreboard to run against the local coding leader, Qwen3.8-27B — the only comparisons are visual side-by-sides (vendor) against DeepSeek V4.1 Flash. One warning: if you grabbed the earlier preview, delete it and download again — the finished weights reuse the same filenames.
Setup notes: Q4_K_M and Q8_0 GGUFs run in LM Studio or another Qwen3.5-capable runtime. Set the output limit to unlimited — replies average around 20K tokens, and an 8K cap cuts the HTML off mid-file. Difficulty: light setup.
Open WebUI v0.13.0 gives your model a computer. Connect Open Terminal once, and every user gets a real desktop, browser and Git in tabs beside the chat — watch it work, take the mouse, or make each step wait for approval. It also carries security fixes and a database migration, so the project's own order stands: back up first.
Immich caught its ML container phoning home. v3.3.6, out yesterday, is a one-fix release: a recent onnxruntime change had turned telemetry on by default, and Immich's ML container was sending anonymous metric data to Microsoft. The fix disables it; the notes apologize and credit the community.
Baidu's 280B base model is a starting line, not a chatbot. ERNIE-5.5-Flash-Base arrived yesterday: 280 billion parameters, 14 billion active per token, trained on 33.5 trillion tokens and stopped before any instruction tuning, for people who want to do that training themselves. FP8 and BF16 weights, Baidu's own license — no EU, UK or South Korea use, and commercial use past $50M in related revenue needs Baidu's written OK. Cluster hardware.
GitHub find: Floci runs AWS on your own box, free. Floci is a local AWS emulator with no account, token or paid gates — point your SDK, CLI or Terraform at localhost:4566, or swap it in for LocalStack. The project's own table claims a 90 MB image and 13 MiB idle against LocalStack Community's 1.0 GB and 143 MiB. 30,530 stars, MIT, latest release October 31.
Boss Fight. Friday, 6 p.m. You run the monthly update on your main services host — new kernel, new container runtime — and reboot into trouble: the reverse proxy will not start on the new runtime, and movie night, the photo sync and the family password vault all sit behind it. Your pre-update snapshot, taken an hour ago, is sitting right there.
Your call:
A. Roll back to the pre-update snapshot now, confirm every service answers, and retry the upgrade on a weekday morning with the failure notes in hand
B. Stay on the new system and hand-start the proxy with a compatibility workaround
C. Re-run the updater and reboot again, hoping a second pass settles it
200 Hero Points for the right call — it is Friday's Boss Fight. Your score is on the outcome page, ready to share.
Simple things should be simple, complex things should be possible.
— Alan Kay
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.
How we source: facts come from primary sources opened for this issue. Benchmarks are labeled (vendor), (independent), or (community), never blended. Deal prices are checked the day they run. This is not buying advice.
Corrections: none today.
Reader-supported: paid subscriptions, labeled sponsors, and affiliate links on some gear and deal items (we may earn a commission). Coverage is never paid for — sponsors and affiliate programs don't pick stories.