AnythingLLM v1.13.0 ships Model Router, and it changes a constraint that has sat under every local AI deployment: you had to pick a side. Run everything locally, or send everything to the cloud. That tradeoff is now optional.
Model Router is a user-defined routing layer that sits in front of every message. When a user sends a query, the system analyzes it and routes it to the model best suited for that specific task. A lightweight local model handles quick questions. A reasoning model handles complex math. A powerful cloud model handles nuanced legal analysis. The user sees one chat interface. The routing is invisible.
You define the rules in two ways. Calculated rules trigger on concrete signals: keywords, token counts, time of day, or image attachments. LLM-classified rules understand intent written in plain English. Both approaches fire instantaneously according to the release notes.
The practical cost argument is straightforward. Route simple queries to cheap or local models. Reserve expensive API calls for messages that actually need them. The release frames this explicitly as a way to save money without sacrificing quality, though no specific cost figures are provided.
One detail worth noting for anyone building multi-turn products: Model Router includes a sticky routing system. Once a conversation thread is on a model, it stays there. You are not bouncing between models on every message within the same thread. That consistency matters for context and coherence in longer conversations.
The hybrid positioning is the real story here. Local models (Ollama, LM Studio, and similar) can be mixed with cloud providers (OpenAI, Anthropic, Google) inside a single conversation. No manual switching is required. The project describes this as the first consumer hybrid AI experience of its kind, and it ships as 100% open source.
A second major feature in this release, Scheduled Jobs, was also announced but the source material is incomplete on its details.
What to do today: If you are building a product on top of AnythingLLM or evaluating open-source AI stacks, Model Router is worth a direct test. The setup documentation walks through both calculated and LLM-classified rule types. The immediate experiment: define one rule that catches your highest-volume, lowest-complexity queries and route them to a local model. Measure what that does to your API spend over a week.