Troubleshooting: the honest failure list

Fervyn provisions its own runtime stack, so most classic local-AI failures (PATH problems, missing Python, wrong Node) can't happen. What's left is a short list. Each entry below says what you'd see, why it happens, and the fix — ordered by how often people actually hit it.

Install and first launch

• "Windows protected your PC" (SmartScreen). The installer isn't code-signed yet — we say so up front. Verify the SHA-256 against your download, then More info → Run anyway. If the file vanished right after downloading, your antivirus quarantined it: restore it from the AV's quarantine and add an exclusion for the install folder before re-running.
• Installer refuses: "needs about 24 GB of free disk space." That check is a floor, not the real first-run budget — it is lower than the download every band actually performs, so passing it is not enough. Budget ~34 GB for the 8 GB config, ~38 GB at 12–16 GB, and ~45 GB at 24 GB and up. Free the space or pick a drive that has it: passing the installer's check with less will stall the download later, and there is no second disk-space check once the fetch starts.
• Install location rules. Install to a local NTFS drive, in a reasonably shallow path (the default is fine). Avoid OneDrive-synced folders and network drives — cloud sync fights the model download and the database.

The one-time download

• Download interrupted / app closed mid-fetch. Relaunch Fervyn. Every download resumes from the last byte — engines, models, all of it. Nothing restarts from zero.
• Download never starts or stalls at zero. Two domains must be reachable the first time: dl.fervyn.app (engines and models — everything ongoing) and registry.npmjs.org (app dependencies, very first launch only). Corporate, school, or parental DNS filters that block either will stall setup. Test by opening https://dl.fervyn.app/Fervyn-Lite-Setup.exe.sha256 in a browser — if that loads, the main download path works; if setup still stalls on the very first launch, allowlist registry.npmjs.org too.
• A step shows "expanding" for a long time. Usually normal, and it can take much longer than you'd expect — the image engine unpacks to about 37,000 small files, and there is no byte counter while it does. On a slow 4-core machine the whole first run can run to several hours, most of it here. Judge it by disk activity, not by the clock: open Task Manager → Performance → Disk for the drive you installed to. If it's busy, it's working — leave it. Treat it as genuinely stuck only if that disk sits flat at zero for 20 minutes or more.

Images

• Renders never start. Almost always an NVIDIA driver older than 580 — the bundled image engine's CUDA 13 build needs r580+. Update at nvidia.com (GeForce drivers), reboot, relaunch. This is the single most common failure on otherwise-working installs.
• First image is slow. The engine loads models on first use; the second image is the honest speed. On the 8 GB config expect a few seconds per panel after warmup.
• "Checkpoint missing" on render. The model download hasn't finished — relaunch to resume it, then retry the render.
• Windows username with non-Latin characters. Handled (the launcher forces UTF-8 for the bundled engines), but if you see encoding errors in an engine window, make sure Fervyn is installed to a plain ASCII path like C:\Fervyn.

Chat

• First reply after a long idle takes a minute. The chat model unloads after idle and reloads on demand. Fervyn pins an efficient context size on 8 GB cards so this stays rare — but the first turn after hours away can be slow once, then it's instant.
• The chat engine pill is red. The first-run download is still running or was interrupted — relaunch to resume. Already run your own Ollama, KoboldCPP, or LM Studio? They're auto-detected; you only need one backend.
• Port conflicts. Fervyn's bundled engines use 11434 (chat), 8188 (image), 8005 (voice), and the app itself uses 3001. If another app owns one of those, Fervyn's copy of that engine steps aside — either close the other app, or point the matching URL in Settings at your own instance.

Voice

• No sound when voice is enabled. The voice engine runs on CPU and starts with the app — give it ~20 seconds after launch, then toggle voice off and on once. Voice is off by default; enable it in Settings.

Reporting a bug that isn't here

Use the feedback form (no account needed) and include: your GPU model + VRAM, NVIDIA driver version, and what the setup screen showed (the models panel lists each component's status — a screenshot of it answers most questions). Fervyn sends nothing automatically: telemetry is opt-in and off by default, so a report from you is the only way we learn something broke.

Everything on this page happens on your machine. After the one-time download, the companion chat path makes zero outbound network calls — the privacy receipt documents the guarantee.