Small, specialised models

When a 6.75M-parameter transformer beats a frontier model: numeric transformers, column tokenisation, synthetic causal data, and training on consumer hardware. Grounded in MNM, which processes 43M+ prior authorisations a year.

Essays

Small, specialised models · 5 min read · with a concept loop

What AI Model Compression Actually Changes Before Release

Distillation and quantization change what an AI model does, not just its size, so test the deployed artifact before trusting any compression claim.

small models and compression inference serving cost and unit economics evaluation and benchmarks

Working on small, specialised models?

Book a conversation