Good morning — it's Monday, November 9. Immich caught its photo server phoning Microsoft and shipped the fix — and the one-variable fix for the rest of your stack. From the bench: Qwopus3.8-27B-Coder-Flash, a coding fine-tune that ties its own base model. Drops: Ollama learns two new model families. Don't keep us a secret: forward this email to a fellow home labber, or share the web version.

Share this issue:  X   WhatsApp   Reddit   Facebook   LinkedIn   Text   Email

Your photo server was phoning Microsoft — Immich shipped the fix

Immich v3.3.6, out November 4, is a one-fix release: a recent ONNX Runtime release had enabled telemetry by default, so Immich's machine-learning container was sending anonymous metric data to Microsoft automatically.

The exposure is wider than one photo app. ONNX Runtime is Microsoft's library for running local models, and faster-whisper, piper-tts and chromadb install it for you. Since version 1.29 in August it has shipped with telemetry on by default on Linux and macOS; in strata→signal's sealed-container test (independent), it tried to reach Microsoft's server 9.2 seconds after starting. The events carry no prompts or outputs, but they do carry a persistent device ID, the path the program launched from, and the names of the models you run.

The fix for everything else is one variable: set ORT_DISABLE_TELEMETRY=1 before the program starts — it is read once, at startup, and setting it later does nothing. In that same test it stopped every attempt. Self-hosting is supposed to end the phone-home; check the dependencies, not just the apps.

Qwopus3.8-27B-Coder-Flash — a coding fine-tune that ties its own base model

Qwopus3.8-27B-Coder-Flash, out this week, fine-tunes Qwen3.8-27B on roughly 200,000 code and reasoning examples, aiming at the agent loop: read, call tools, edit, test, recover, deliver. About 56 GB in BF16; the author's GGUF repo is the homelab route in.

The scoreboard — higher is better throughout. Verdict first: on the one independent head-to-head, Qwopus ties the base Qwen3.8-27B; the rest is the author's own local runs.
Real-bug coding test, 25 bugs (independent), Samuel Batista's test: Qwopus 18.5 fixed, Qwen3.8-27B 18.0 — ahead by 0.5, inside run-to-run noise.
MMLU-Pro subset, 1,500 questions (vendor): Qwopus 93.87% — run in the author's harness only; no comparable base score exists.
Seeded agent tasks (vendor): Qwopus 9 of 9 runs — the author's own three-task set, which the card itself calls small and targeted.

The evidence line: Batista screened 15 Qwen3.8 variants and found Qwopus worse on facts and long conversations, level on coding; the card reports no SWE-bench or Terminal-Bench at all. Verdict: worth a weekend spin, no proven gain over the base yet. Setup notes: GGUFs from the author's repo for llama.cpp and LM Studio; vLLM serves the full checkpoint. Difficulty: light setup.

Ollama v0.40.15 adds MiMo and a decision-model family. Xiaomi's mimo-v2.6, a 9B model built for code and agent work, now runs with one command. The new AutoTrust models return choices, yes/no probabilities and ratings instead of prose — in 9b, 35b and 27b-vl sizes, the larger two reading images.

An htop for what's inside your GPU. Scott Hanselman's gputaskman (1,139 stars, MIT, seven releases since November 3) shows per-process VRAM and, for llama.cpp, Ollama and LM Studio, splits a loaded model into weights, KV cache and compute buffers. Windows, Linux and Apple Silicon; the repo calls it early and experimental, breakdowns labeled estimates.

GitHub find: Agent Autoloop lets your CLI agent work while you're away. KKKKhazix/agent-autoloop (4,158 stars, MIT, 14 releases since November 3) matches events — a cron time, a new GitHub Issue, a webhook — to a command-line agent. The neat part: when the agent hits a call it shouldn't make alone, it exits, sends you an approval card, and burns no tokens while it waits. One catch: the docs are mostly Chinese, with an English README in the repo.

Monday, 7 a.m., and the overnight image pull went wrong: your reverse proxy container is crash-looping, every service behind it is down, and the house is waking up to dead bookmarks and a photo backup that cannot start. Coffee is not yet a factor.

Your call:
A. Read the container's logs, find the first error, and fix it — or roll back to the previous image — before starting it once
B. Restart the whole Docker host and hope the loop clears on a fresh boot
C. Delete the container and its volumes and recreate it fresh — clean slate, no bad state

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.