Each background workload can use a different model than the bot. Leave blank to inherit the active LLM above. Subtitle / Topic analysis want a conversational model — try openrouter / google/gemma-3-27b-it:free or googleai / gemini-2.5-flash.
Vision captioning needs a multimodal model — try googleai / gemini-2.5-flash (cheap cloud) or openrouter / anthropic/claude-sonnet-4.6 (premium); local Qwen2.5-VL via LM Studio also works.

Per-task model picks. Live cheatsheet ranking + user pins. Owner can pin any alternative (requires DASHBOARD_OWNER_TOKEN); non-owners see read-only.