Good morning — it's Saturday, November 7. Cisco's C1300-24XTS puts 24 10GbE ports in one rack unit, half copper, half fiber. From the bench: TwIL-8B, an 8.8B reasoner. Drops: Ceph Umbrella in Proxmox preview, and Stirling-PDF v3.3.0. Forward this to a fellow home labber.

Share this issue:  X   WhatsApp   Reddit   Facebook   LinkedIn   Text   Email

Twenty-four 10GbE ports, half copper, half fiber — and line rate on all of them

Cisco's Catalyst C1300-24XTS, reviewed by ServeTheHome yesterday, splits its 24 10GbE ports evenly: twelve 10GBase-T copper and twelve SFP+ cages — not combo jacks. The 1U box adds a separate 1GbE management port, and ServeTheHome has run it for months without issue.

In ServeTheHome's Keysight tests (independent), it hit line rate across all 24 ports with zero frame loss: just under 183Gbps at hard 64-byte frames, about 227Gbps on IMIX traffic. Power is the trade: 19.6W idle, but one SFP+ module took the meter to 28.0W, at 38 dBA. And there is a single power supply — redundant power means an external transfer-switch PDU.

Why now: new local AI boxes — GB10 machines, Ryzen AI Max+ systems, GB300 stations — arrive with 10GBase-T ports, while storage still speaks fiber. A switch that takes both — with firmware downloadable without a Cisco login — is the rack's missing middle: Cisco-grade hardware, no support contract attached.

TwIL-8B — an 8.8B reasoner that tops its own class, not the local leader

TwIL-8B, out this week from webAI, is a reasoning model for math, code and instruction following, post-trained from IBM's Granite 4.2 8B. It ships in two tempers: balanced, the default, and verbose. Inside its own class it leads (vendor): MATH-500 98.2, best in its comparison group. The card is honest — it states that this is not a knowledge model, and that SWE-bench was never evaluated.

Against the local leader, Qwen3.8-27B, the leader wins every shared row — higher is better, all scores (vendor) from each model's own card:
GPQA Diamond: TwIL-8B 70.2 · Qwen3.8-27B 89.2 — the leader by 19.0
IFBench: TwIL-8B 74.4 (verbose) · Qwen3.8-27B 79.5 — the leader by 5.1
HLE: TwIL-8B 12.6 (verbose) · Qwen3.8-27B 30.8 — the leader by 18.2, though the two sides score it differently
One row is not lined up: TwIL's card says its LiveCodeBench run's version is unrecorded, so its 81.6 cannot fairly sit beside the leader's v6 score. Nobody independent has replicated either side yet.

Setup notes: the 4.98 GiB 4-bit GGUF runs in llama.cpp on modest hardware. Give it a large generation budget — balanced averages about 19,500 tokens per AIME 2026 problem, and a short cap truncates the reasoning. Scores were measured on the full-precision weights, so expect small deviations. Difficulty: light setup.

Ceph Umbrella lands in Proxmox as a preview — and Squid is nearly done. Proxmox's announcement puts Ceph 21.2 Umbrella in the no-subscription and test repositories as a technology preview; the enterprise repository waits until Proxmox calls it production-ready. The push inside the post: Ceph 19.2 Squid has one last update left — upgrade sooner rather than later.

Stirling-PDF v3.3.0 is a fork in its converter road. The release notes upgrade the beta Office Convert and let you pick it as the converter inside other tools; the roadmap turns it on by default next week and, if it holds, removes LibreOffice the week after — so conversion bugs are worth reporting now. Also in: better redaction, lower memory on big files, and security fixes the project advises installing.

GitHub find: Kanzo is a kanban board you share with your coding agents. Kanzo is a browser board and an MCP server in one 4.7 MB program: you steer while your agents take tasks, move cards and comment over MCP, and the board syncs between machines through a private git repo. No account, no server. 1,074 stars, AGPL-3.0, updated yesterday.

Saturday morning is drive-swap day on your NAS, and you do the right thing first: a test restore of the photo-library VM from last night's backup. It fails. The backup job has reported success 41 nights straight — and the rebuild you were about to start would have leaned on those backups.

Your call:
A. Call off the swap, test-restore the older generations and the offsite copy, and fix the backup chain before touching the array
B. Swap the drive, but keep the pulled drive untouched on the shelf as your fallback while you sort the backups out
C. Re-run the backup job, watch it report success, and start the rebuild

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.