A financial analyst asks a general-purpose chatbot about the mechanics of a SOFR-linked repo transaction and gets an answer that's plausible-sounding but subtly wrong about a market convention that only shows up in specialized terminals and internal desk jargon. Bloomberg's response to exactly this gap was BloombergGPT, a model trained substantially on the company's own decades of financial data alongside general text — a deliberate bet that depth in one domain would outperform breadth across all of them for the queries its own terminal users actually ask.
A general-purpose model optimizes for competence across an enormous range of possible queries. A domain model trades that breadth for depth in exactly the area where the decisions happen, catching context-specific nuance a universal model, however capable, was never trained to weight heavily because it was drowned out by everything else in a broader training mix.
Apple makes a version of the same bet at a different layer: Apple Intelligence runs smaller models on-device for personal-context tasks, escalating to a larger model in Apple's Private Cloud Compute only when the task genuinely requires it. Microsoft's Phi family of small language models is built on the same premise — that a model deliberately kept small and focused can outperform a much larger general one on the narrow tasks it's actually asked to do, while running somewhere a giant model can't.
Organizations with genuinely idiosyncratic context — specialized regulatory environments, internal jargon that means something different elsewhere, edge cases a universal model would treat as noise — see the clearest benefit from this approach, because their edge cases are precisely the material a general model was never given enough weight to learn.
Local models also carry a privacy and latency advantage that has nothing to do with raw capability: sensitive data that never has to leave a device or a facility, and response times that don't depend on a round trip to a distant data center. Apple's entire Private Cloud Compute architecture is built around minimizing exactly how often that round trip has to happen.
The likely shape of the next few years is not universal models replacing local ones or vice versa, but a layered stack — a general model for broad, unfamiliar queries, and smaller local or domain models for the narrow, high-frequency decisions where local knowledge beats general capability. Bigger stops being a synonym for better the moment the task depends on knowing something specific that only the smaller, closer model was ever taught.