Route work with models and profiles
A provider tells Lotti where to connect. A model records one provider model ID and its capabilities. An inference profile then assigns those models to the jobs Lotti understands. This keeps “which model handles audio?” out of every individual task and automation rule.
Review saved models
Open Settings → AI Settings → Models. The list groups models by provider and exposes capability filters for text, vision, audio, and reasoning.
The Concurrent agent wakes control at the top of AI Settings limits how many different agents may run inference at once. A higher value can improve throughput, but it also increases simultaneous provider requests and local device load. Keep it conservative when a provider has strict rate limits or a local model server has limited memory.
Add or edit a model
Select a model card, or use + to add one. A model definition includes:
- the provider connection;
- a friendly display name and optional description;
- the provider's exact model ID;
- the maximum completion-token limit;
- input and output modalities; and
- capability flags such as reasoning or function calling.
Use the exact identifier expected by the provider. The display name is only for people; changing it does not change what the provider runs. Do not claim image, audio, tool, or reasoning support unless that model and endpoint actually provide it.
Build an inference profile
Open Settings → AI Settings → Inference Profiles. Each profile is a reusable routing policy. The overview shows which model serves each configured slot.
Edit a profile to assign models for the capabilities you use:
- Thinking is the required general text model.
- Thinking (High-End) can handle work that explicitly requests the more capable reasoning path.
- Image Recognition analyzes images.
- Transcription converts audio to text.
- Image Generation creates task artwork and other generated images. The current built-in image-generation path uses a configured remote provider, so choose that provider and the references you send with the same care as any other cloud inference route.
Desktop Only hides a profile on mobile platforms, which is useful when it depends on a model reachable only from a workstation. Device pinning is for automated synced-audio processing: pin only when the selected device advertises the providers needed by the profile. With no pinned device, synced audio is not automatically transcribed by that profile.
Choose a model without memorizing IDs
Model pickers start with the provider and show only compatible saved models. The picker in a profile slot, for example, will not offer an audio-only model for text reasoning.
Set a category default
A category can choose a default inference profile under Settings → Definitions → Categories. New work in that category can then inherit a coherent routing policy. Select the profile whose privacy and capability mix matches the category rather than choosing solely by speed.
Legacy profile route
The older standalone Inference Profiles route remains available for compatibility and edits the same stored profile records. It exposes internal model IDs and provides less dependency context than AI Settings, so prefer the current AI Settings → Inference Profiles page for normal maintenance.
Continue with AI skills to see how the selected profile becomes an action on a task or entry.