# Why are tiny models (<50M parameters) or swarms of specialised micro-models so rarely deployed in production? — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/b20ff9f608567a1f

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-09-13 18:20 ingested 2026-09-14 02:38

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1wffbrg/why_are_tiny_models_50m_parameters_or_swarms_of/)

## retained text



```
I have been thinking about why we do not see more tiny, specialised models in production. It feels like it would be so much more efficient to use small, task-specific ones for certain things, but we always seem to end up with one massive model doing everything.

Is it just because our tools and inference engines are built for big models, or is it just easier to prompt a generalist than to do the hard work of training a specialist? I would love to hear what people working at scale are seeing.
```



document fetched 2026-09-14 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1wffbrg/why_are_tiny_models_50m_parameters_or_swarms_of
