[tutorial] Improve Local LLM Performance with Dedicated Task Models

[taylorwilsdon]

Have seen multiple reddit posts and github issues recently where new users had the "Current Model" selected for the default tasks, and were running a model that barely fit on their GPU so autocomplete took the performance from sluggish to unusable. Thought it might be beneficial to put together a quick doc on how to configure these optimally and reduce confusion from users. Happy to make any changes if you'd like @tjbck - wasn't sure if I should list providers by name (ie openrouter) or if there are any recommended API providers you already work with.

This pull request adds updated documentation aimed at improving the performance of Open-WebUI when running local LLMs. The new guide explains why performance issues often occur—primarily because of multiple background automation tasks—and provides clear, step-by-step instructions to optimize task model performance.

• A detailed explanation of the background tasks (title generation, tag generation, autocomplete, and search query generation) that can burden local resources.
• Clear instructions to configure a dedicated, lightweight task model (e.g., Llama 3.2 3B or Qwen2.5 3B) via the Admin Panel.
• Recommendations for connecting to hosted API endpoints (such as OpenRouter or GLHF.chat) to offload processing and enhance responsiveness.
• Tips on disabling unnecessary automation features for users who prioritize core chat functionality.

These enhancements ensure that users can achieve a faster and smoother experience with Open-WebUI by optimizing their configuration based on their hardware capabilities or network access.

[taylorwilsdon]

Image won't render in the github preview but it's correctly included based on existing examples in the repo for docusaurus to host and display inline

[tjbck]

Thanks!