Specialists, composed on demand.
An expert model is small, narrow, and good at one thing. The mesh holds many of them and lets the router decide which ones participate in any given request.
The case for many small models.
A 7B specialist that is genuinely good at SQL will outperform a frontier generalist on SQL queries — at a fraction of the cost. The same is true of code, of OCR, of intent classification, of structured extraction. The frontier of real-world utility is full of these wins, and almost none of them are captured by single-model deployments.
We think the right abstraction is a capability, not a model. An expert is a typed function the mesh can call. Some are large, some are tiny, some are not even neural. The router does not care; it cares about whether the capability is available and whether it is the right one for the subtask.
Domains we are exploring.
Language
Translation, summarization, intent, structured extraction — small models, large quality on narrow tasks.
Code
Synthesis, refactoring, type-aware reasoning. Code is one of the cleanest domains for specialization.
Vision
Object understanding, OCR, diagram parsing, document layout — modalities routed only when present.
Documents
Long-context comprehension over PDFs, contracts, and structured corpora with retrieval grounding.
Structured Data
SQL, JSON, tabular reasoning. The class of problems where general LLMs are routinely outperformed.
Math & Logic
Symbolic reasoning, formal verification, theorem-style decomposition — domain experts beat generalists.
What every expert in the mesh has in common.
Capability-addressed
Experts are referenced by what they can do, not which weights they are. The router selects on capability, not vendor.
Composable
An expert is a black box with typed inputs and outputs. The mesh stitches them together; the model does not need to know.
Dynamically utilized
Most experts sit idle on any given request. The mesh only spends inference where the router thinks it will help.
See how experts coordinate across modalities.
Vision, text, audio, and structured data experts share a context layer. Read the multimodal page for how that works.