Troubleshooting

Common problems and what to do about them. If none of these fit, export a diagnostics bundle and ask in the community.

”Bedrova can’t be opened” on first launch

That’s macOS Gatekeeper reacting to a fresh download. Right-click the app and choose Open, then confirm. You only do this once. Shipping builds are signed and notarized by ideius, so a release .dmg or the Homebrew cask should open without this step.

A model won’t load, or “out of memory”

This is the fit-check doing its job — it stops a load that wouldn’t fit instead of letting the app crash. Any of these resolves it:

• Free memory by unloading another model (or letting an idle one auto-evict).
• Pick a smaller quant — a 4-bit build of the same model can cut peak memory by more than half.
• Shorten the context length in the model’s profile; KV cache grows with context.
• Accept the suggested alternative the fit-check offers.

If you genuinely want to push past the warning, the guardrail mode in Settings → Model Loading Guardrails controls how strict the check is — see Memory & loading.

Generation feels slow

Speed depends on the model, quant, and context length, in that order of impact.

• Try a smaller or more heavily-quantized model.
• Shorten the prompt and context; long contexts slow the prefill.
• Watch Settings → Hardware during a long run for thermal or memory-pressure warnings — a hot Mac throttles.
• Where supported, turn on speculative decoding with a draft model. See Performance.

A request returns “overloaded” or gets queued

A workload lane is saturated. The response is retryable by design — retry, or reduce how many requests you send in parallel (Settings → Server/API → max parallel requests). Background on the model is in Parallel requests & scheduling.

Updating or rolling back

App updates arrive through Sparkle (Settings → General), signed and disclosed. Runtime and provider packs can be pinned to a version or rolled back individually in Settings → Runtime — useful if a new pack regresses a model you depend on. See Runtime & engines.

Models on an external drive

Point the model directory at your drive in Settings → General. Bedrova uses security-scoped bookmarks so access survives restarts, detects disconnects, and offers to repair paths when the drive comes back. Throughput on the drive affects load time. More in Data & file locations.

A downloaded model is quarantined

New downloads stay quarantined until hash, manifest, and license checks pass, and models that need trust_remote_code are held back because that’s off by default. See Model trust & verification for how to review and clear one.

Screenshot: the diagnostics bundle export — added at launch.

Still stuck

Export a diagnostics bundle (Settings → Diagnostics — it redacts prompts, files, and secrets) and ask in the community. Diagnostics & support bundles explains exactly what the bundle contains.